Trying to Build a Digital Brain for the University

1. The Dream: a Digital Brain

Navigating a university website like https://univlora.edu.al often feels like digital archaeology. Vital information — admissions criteria, academic regulations, and administrative mandates — is frequently fragmented across deep directories, nested subpages, and PDF documents. My goal was to synthesize this chaos into an intelligent system, a "digital brain" that can answer questions related to UV, based on the information and documents that are already published on the website.

The technical term for such a system is: Retrieval-Augmented Generation (RAG). It is an architectural framework designed to improve the accuracy and reliability of Large Language Models (LLMs) by giving them access to data outside of their initial training set.

About the RAG system

Think of an LLM as a brilliant student who has read every book in the world but graduated three years ago. If you ask them about yesterday’s news, they might guess or "hallucinate." RAG is like giving that student an open-book exam with access to a modern library or a private database. In our case, the LLM would be able to give accurate answers based on the information provided by the university website.

A RAG system operates through a multi-step pipeline that bridges the gap between raw data and a generated response:

Retrieval: When a user asks a question, the system searches a specific "knowledge base" (like your company’s internal documents, a wiki, or a collection of PDFs) for relevant information.
Augmentation: The system takes the user’s original query and "augments" it by attaching the relevant snippets found during the search.
Generation: This combined package (Query + Context) is sent to the LLM. The model uses the provided context to write a response that is grounded in facts.

To build a RAG system, several components must work in sync:

Vector Database: Because LLMs don’t "read" text like humans, data is converted into numerical representations called embeddings. These are stored in a Vector Database (such as Milvus, Pinecone, or Weaviate), which allows the system to find information based on mathematical similarity rather than just keyword matching.
The Embedding Model: This is a specific model (like OpenAI’s text-embedding-3 or HuggingFace models) that translates human language into vectors. It ensures that "The server is down" and "The machine isn’t responding" are recognized as having the same meaning.
Orchestration Layer: Frameworks like LangChain or LlamaIndex act as the "glue" that connects the database, the search logic, and the LLM into a single automated workflow.

Why Use RAG?

Reduces Hallucinations: The model is forced to cite its sources and stick to the provided text.
Up-to-Date Information: You don’t need to retrain the model to teach it new facts; you just update the documents in your database.
Data Privacy: You can feed the model sensitive or proprietary data (like internal rules or regulations) without that data becoming part of the public model’s permanent memory.
Cost-Effective: Fine-tuning an LLM is expensive and time-consuming. RAG provides specialized knowledge at a fraction of the cost.

Trying to build a RAG system for the UV seemed like a straightforward weekend project. The plan was elegant: gather all the university regulations, guidelines, and curriculum documents, feed them into some AI tool, and embed a sleek chatbot on our website so that students and staff could get instant answers in Albanian.

But as any programer will tell you, things are never as easy as they seem on the first sight. From the domain boundaries of cloud-hosted platforms to the brutal reality of running optical character recognition (OCR) on scanned documents using GPU-less servers, my journey was a series of technical battles.

2. The Easy Way: NotebookLM

Initially, I realized that a simple and highly accurate RAG system can be built in minutes using Google’s NotebookLM. By creating a notebook and uploading the PDF guidelines, university rules, curriculum documents etc. we immediately get a highly grounded, conversational assistant, that cites its sources perfectly. The language of the documents can be Albanian, and the questions and answers can be in Albanian of course (as well as in English or any other language).

I have noticed that the Albanian of Gemini/NotebookLM is perfect, even better than mine.

Then, this notebook can be shared with other people, such that only a small group of people have full access (for keeping the uploaded documents up to date), and the rest can only ask questions and get answers based on those documents, but cannot modify the documents or other aspects of the notebook.

How to share a notebook

To collaborate with others, look for the Share button (usually a person icon with a plus sign) in the top-right corner of your notebook:
- Individuals & Groups: You can invite specific people or an entire Google Group by typing their email addresses.
- Permissions: You can set collaborators as either Viewers or Editors.
  - Editors can add/remove sources, create notes, and chat with the AI.
  - Viewers can read the sources and notes and interact with the chat, but they cannot modify the notebook’s structure.
The "Chat View" (the sleeker interface)

When you share a notebook, you don’t have to give people the full Editor view with all your notes and source lists. You can specifically select "Copy link to chat view".
- This provides a cleaner, simplified interface that looks more like a dedicated Q&A web page.
- Users can ask questions immediately, and the AI will answer using only your curated documents.

Figure 1. The notebook with all the sources

Figure 2. The chat view that can be used by anyone

A problem with this approach is that the notebook can be shared only with the members of the organization, because UV is using the Google Workspace Education Fundamentals tier (which is free). Sharing it with anonymous people who might have any questions about UV is not possible.

Another problem is that it is not possible to embed it as a nice chatbot on a web page (for example on the website of UV).

Using a personal Gmail account

By the way, if we moved the notebook project to a personal Gmail account, it would be possible to generate a "Public Link" (Chat View), which allows anyone to ask questions to the provided documents, without seeing the document list or other details of the notebook.

However, if we try to embed the Chat View link inside a standard HTML <iframe> on a public university portal, the browser blocks it and displays a "Refused to connect" error.

This is because most Google Workspace core services (including NotebookLM) use a security header called X-Frame-Options: SAMEORIGIN. This is a piece of code Google sends to the browser that effectively says: "Only show this content if it is being hosted on a google.com website."

Some other Notebook limits

Last time I checked, Notebook had these limits, but they may change with time. They were the same for the Google Workspace Education Fundamentals tier and for a Personal Account.

Source & Notebook Limits

These define how much information you can "feed" the AI in a single project.
- Sources per Notebook: Up to 50 sources. (If your school upgrades to Education Plus, this jumps to 300 sources).
- Word Count per Source: Up to 500,000 words per individual document.
- File Size per Source: Up to 200MB for local file uploads (PDF, text, or markdown).
- Total Notebooks: You can create up to 100 notebooks in your account.
Daily Usage Limits

These limits reset every 24 hours and apply to how you interact with the notebook.
- Chat Queries: 50 questions per day. (This is the most common limit users hit during heavy research).
- Audio/Video Overviews: 3 generations per day. These are the podcast-style or video explainer summaries.
- Study Tools: Up to 10 quizzes or sets of flashcards per day.

3. Self-Hosted RAG Platforms

To gain complete control and embed the chat assistant directly on our website, I turned to self-hosted open-source wrappers. Initially I tried AnythingLLM (https://llm.univlora.al/), and later I also tried:

These platforms are outstanding — they provide clean multi-user interfaces, workspace organization, and excellent document management tools. Most of them also allow you to embed an AI chat assistent on your website (for example AnythingLLM does).

But these platforms come with a major catch for self-hosters: they are completely empty engines.

None of them provides AI models out of the box. They act as the orchestration and database layer, but you must supply:

An embedding model to vectorize the documents (such as nomic-embed-text or bge-m3).
An LLM model to synthesize the retrieved context and formulate the response.
The data or documents from your website, because usually they are not able to crawl, scrape and process a whole website automatically.

The AI models required by them can be provided by service providers like Gemini, Claude, Anthropic etc. (using API keys, subscriptions, etc. which cost money).

But they can also be locally installed models (self-hosted), typically managed and served via Ollama.

The subscription options were not feasible for me, even just for testing, because I don’t have any budget for these project. So, I focused my efforts only on exploring and evaluating the self-hosted options/solutions.

4. The Art of Web Crawling

With AnythingLLM ready, I needed to feed it the information from the website. I attempted to crawl the University of Vlora’s website (https://univlora.edu.al) to download all published pages and PDFs.

I found some nice crawling tools:

siteone-crawler (both cli and desktop version)
spider_cli (only cli)

Here are a couple of other options that I considered trying, but gave up because they seemed a bit too complex for the time being:

Firecrawl (self-hosted)
Crawl4AI (self-hosted)

I eventually used spider_cli to crawl the website. Here is a basic script:

Script 1: spider.sh

#!/bin/bash

url='https://univlora.edu.al/'
output_dir='./scraped_site'
agent="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36"

mkdir -p $output_dir

spider \
    --url "$url" \
    --subdomains \
    --return-format markdown \
    --headless \
    --block-images \
    --stealth \
    --delay 300 \
    --agent "$agent" \
    download \
    --target-destination $output_dir

However this crawling scripts ran into several obstacles. For example, there are some PDF and office docs on the website, that are available for view or download, but they are shared (publicly) from google drive. The script is missing them for two reasons:

External Domain Boundaries: By default, spider_cli restricts crawling to the target domain (and its subdomains). Links pointing to external domains like drive.google.com or docs.google.com are ignored.
Google Drive Link Mechanics: Google Drive share links (e.g. https://drive.google.com/file/d/…/view]) point to an HTML/JavaScript viewer web application, not a raw binary file (.pdf or .docx). Even if spider_cli followed external links, spider download would only save the wrapper HTML page of Google Drive’s viewer rather than downloading the actual file.

Another obstacle is that the spider command is failing to download PDF docs from the site, maybe because they are displayed on an iframe inside the webpage. spider_cli often skips <iframe> targets because it treats iframe source URLs as embedded frame assets rather than standard navigational links. This is a severe problem because it turns out that most of the information on the UV site is published as PDF docs.

The most reliable way to handle this in a Bash workflow is a pipeline like this:

Use spider_cli to crawl and download the site’s HTML pages.
Extract all Google Drive URLs from the scraped content and download the underlying files using gdown (a CLI tool designed to handle Google Drive sharing links and direct file extraction; pipx install gdown).
Parse the scraped HTML files for <iframe> src attributes and direct .pdf references, resolve relative paths against the domain, and download them using curl or wget.

The crawling scripts

With some help from Gemini, I crafted this crawling script that downloads all the HTML pages, as well as Google Drive PDFs and other docs from the website:

Script 2: spider.sh

#!/bin/bash

url='https://univlora.edu.al/'
output_dir='./scraped_site'
gdrive_dir="./gdrive_files/"
agent="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36"

mkdir -p $output_dir
mkdir -p $gdrive_dir

echo "=== Step 1: Crawling and downloading website with spider_cli ==="
spider \
    --url "$url" \
    --subdomains \
    --return-format markdown \
    --headless \
    --block-images \
    --stealth \
    --delay 300 \
    --agent "$agent" \
    download \
    --target-destination $output_dir

echo "=== Step 2 & 3: Extracting IDs and Downloading Files ==="

# 1. Extract File IDs from drive.google.com/file/d/<ID>
drive_ids=$(grep -Eohr 'drive\.google\.com/file/d/[a-zA-Z0-9_-]+' "$output_dir" | sed 's|.*file/d/||' | sort -u)

# 2. Extract File IDs from docs.google.com/document/d/<ID> (Native Google Docs)
doc_ids=$(grep -Eohr 'docs\.google\.com/document/d/[a-zA-Z0-9_-]+' "$output_dir" | sed 's|.*document/d/||' | sort -u)

if [ -z "$drive_ids" ] && [ -z "$doc_ids" ]; then
    echo "No Google Drive or Docs links were found."
fi

# Download standard Google Drive files (PDFs, Office docs, ZIPs, etc.)
for id in $drive_ids; do
    echo "Downloading Drive File ID: $id"
    gdown "$id" -O "$gdrive_dir/" || echo "Failed to download $id"
done

# Export native Google Docs as PDFs
for id in $doc_ids; do
    echo "Exporting Google Doc ID as PDF: $id"
    curl -sL "https://docs.google.com/document/d/$id/export?format=pdf" -o "$gdrive_dir/$id.pdf"
done

echo "=== Download process complete ==="

The script get_pdfs.sh extracts the URLs of the embedded or iframe PDFs from the HTML files, and downloads them:

Script 3: get_pdfs.sh

#!/bin/bash

url='https://univlora.edu.al/'
output_dir='./scraped_site'
gdrive_dir="./gdrive_files/"
agent="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36"

echo "=== Step 4: Extracting and Downloading PDFs (Preserving Site Structure) ==="

pdf_dir="./embedded_pdfs"
base_url="${url%/}"
domain=$(echo "$base_url" | sed -E 's|^https?://||' | cut -d'/' -f1)

# Extract relative and absolute PDF/iframe URLs
raw_pdf_links=$(grep -Eohr 'https?://[^\t\n"'\'' >)]+\.pdf(\?[^\t\n"'\'' >)]*)?|(src|href)=["'\'']/?[^"'\'' >]+\.pdf|<iframe[^>]+src=["'\''][^"'\'' ]+' "$output_dir" | \
  sed -E 's/(src|href)=["'\'']//; s/<iframe[^>]+src=["'\'']//' | sort -u)

if [ -z "$raw_pdf_links" ]; then
    echo "No embedded or iframe PDFs found."
else
    for item in $raw_pdf_links; do
        # Skip external domains (handled separately by gdown or ignored)
        if [[ "$item" =~ ^https?:// ]] && [[ ! "$item" =~ $domain ]]; then
            continue
        fi

        # 1. Build complete absolute URL for curl
        if [[ "$item" =~ ^https?:// ]]; then
            pdf_url="$item"
        else
            clean_item="${item#/}"
            pdf_url="$base_url/$clean_item"
        fi

        # 2. Extract path portion (e.g. https://example.org/docs/file.pdf -> /docs/file.pdf)
        url_path=$(echo "$pdf_url" | sed -E 's|^https?://[^/]+||')
        clean_path=$(echo "$url_path" | cut -d'?' -f1) # Strip query parameters (?v=123)

        # 3. Map path directly onto $output_dir
        target_file="${output_dir%/}${clean_path}"
        target_dir=$(dirname "$target_file")

        # 4. Recreate directory tree and download
        mkdir -p "$target_dir"
        echo "Saving: $pdf_url"
        echo "   ---> $target_file"
        curl -sL -A "$agent" "$pdf_url" -o "$target_file" || echo "Failed to download $pdf_url"
    done
fi

5. PDF OCR and Preprocessing

Once I successfully downloaded the website, including all the PDFs, the next step was document processing, in order to create a dataset and upload/injest it to AnythingLLM. The HTML and PDF formats are not quite suitable for AnythingLLM, but plain text or markdown (.txt or .md) are perfect.

How to dump the data into the workspace folder of AnythingLLM

Assuming that we have installed and setup AnythingLLM properly, and we have created a workspace for this project, we can create a dataset for this workspace like this:

Convert the downloaded pages and PDF docs to markdown (.md).
Move those files directly into the AnythingLLM storage directory, on the host system: <anythingllm_volume>/_data/documents/
Open the AnythingLLM UI, navigate to the Document Manager, and we will see the entire file tree ready to be selected and processed into the vector pool.

Converting HTML pages to markdown is rather easy, even a simple tool like pandoc does it well. There are also some more advanced tools that remove the header, footer, and the stuff that repeats on every page, in order to keep only the most important information of the page.

However, it turns out that most of the information on the UV website is published as PDF docs, the information of HTML pages is very slim. So, the most important task is actually converting PDFs to markdown (or plain text).

A simple tool for extracting the text from a PDF file is pdftotext:

apt install poppler-utils    # provides pdftotext
pdftotext file1.pdf

But here we hit another problem: most of the PDFs published online are scanned, image-only documents. Standard PDF parsers (like pdftotext) extract zero text from these files. We need some tool that can apply image processing, and optical character recognition (OCR) on them, in order to extract the text from them.

OCRmyPDF is such a tool, that is easy to install and use:

# install packages
apt install \
    ocrmypdf \
    tesseract-ocr \
    tesseract-ocr-sqi \
    unpaper

# process vba7-2025.pdf and generate vba7-2025.ocr.pdf
ocrmypdf -l sqi+eng \
    --force-ocr \
    --deskew \
    --clean \
    --rotate-pages \
    --oversample 300 \
    --tesseract-pagesegmode 1 \
    vba7-2025.pdf \
    vba7-2025.ocr.pdf

# extract the text to stdout
pdftotext vba7-2025.ocr.pdf -

An example

The original PDF document:

The text extracted from OCR:

REPUBLIKA

E

SHQIPËRISË

UNIVERSITETI “ISMAIL QEMALI” VLORË
BORDI I ADMINISTRIMIT

Vlorë mé, 94 01.2025

VENDIM

Nr..1, datë 09.01.2025
PËR PAGESËN E ANËTARËVE TË GRUPEVE TË PUNËS PËR SHKRIM
PROJEKTESH

Në mbështetje të nenit 49, pika 1, gërma ë, të Ligjit Nr. 80/2015 “Për arsimin e lartë dhe
kërkimin shkencor në institucionet e arsimit të lartë në Republikën e Shqipërisë”, nenit 25, të
Statutit të Universitetit “Ismail Qemali”, Vlorë, Udhëzimit nr. 29, datë 10.09.2018 “Për
veprimtarinë dhe ngarkesën mësimore të personelit akademik në institucionet e arsimit të

lartë”, si dhe referuar propozimit të Administratorit ardhur me shkresë 3122, prot, datë
26.12.2024, Bordi i Administrimit, në mbledhjen e datës 09.01.2025,
VENDOSI :

1.

Të miratojë pagesën e grupeve të punës sipas materialit bashkëlidhur dhe pjesë e
këtij vendimi.

2. Ngarkohet për zbatimin e ketij vendimi Administratori dhe Drejtoria e Shërbimeve

Mbështetëse të UV- së.

L

Ky vendim hyn né fuqi menjéheré.

KRYETAR

OPUSAE mS

te

thf

Adresa /Adress: Bulevardi Vlore-Skele
Tel. Fax ++355222288

V

er?

Web-site: www.univlora.edu.al

e-mail: info@univlora.edu.al

REPUBLIKA E SHQIPËRISË

UNIVERSITETI “ISMAIL QEMALI” VLORË
ADMINISTRATOR

Vlorë, më 16. 11 .2024

Nr. KA 3 L...Prot
Lënda:

Përciellje për miratimin e efektit financiar për shkrim/hartim projektesh.

BORDIT TË ADMINISTRIMIT TË UNIVERSITETIT "ISMAIL QEMALI" VLORË
Z. SOTIR NIKOLLA

Në zbatim të Ligjit nr. 80/2015 “Për Arsimin e Lartë dhe Kërkimin Shkencor në Institucionet e Arsimit
të Lartë në Republikën e Shqiperisë”, Udhëzimit nr. 29 datë 10.09.2018 për “Veprimtarinë dhe
ngarkesën mësimore të personelit akademik në IAL”, ligjit nr. 9936 datë 26.06.2008 “Për menaxhimin e
sistemit buxhetor në Republikën e Shqipërisë”, i ndryshuar, ligjit nr. 10296 datë 08.07.2010 “Për
menaxhimin financiar dhe kontrollin” ligjit nr. 97723 “Për buxhetin e vitit 2024”, vendimit të Bordit të
Administrimit nr. 8 datë 23.02.2024 “Për miratimin e buxhetit të Uinversitetit “Ismail Qemali” Vlorë për
vitin 2024”, Vendimin e Senatit Akademik nr. 17 datë 06.05.2021, vendimit të Bordit të Administrimit
nr. 45 datë 05.12.2023 “Për miratimine rregullores financiare të Universiteti “Ismail Qemali” Vlorë”, si
dhe shkresave të Drejtorisë së Sigurimit të Brendshëm të Cilësisë dhe Projekteve shkresa nr. 2989/1
prot., datë 20.12.2024 “Relacion mbi shkresën me nr. 1978/2 prot., datë 11.12.2024” efekti financiar për
grupin e punës të cilat përmbushin çdo detyrim të vendimit tëSenatit Akademik nr. 17 datë 06.05.2021
dhe rregullores financiare miratuar me vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 është
si më poshtë :
Pozicioni)

Nr
1
2
3
4
4

Emer Mbiemer
Fjoralba Velaj
Zamira Sinaj
Rezarta Brokaj
Migena Petanaj
Totali

Projekti

titulli”

grada
Prof.As
Prof.As
Dr
Dr

WIKADIH
(Aplikim
viti 20232024)

Viera/ore
1,295
1,295
940
940

Oret projekt
70
20
30
20
140

Vlera

Oret e punes

Totale/Leke
90,650
25,900
28,200
18,800
163,550

se pagueshme
70
20
30
20
140

Bazuar në shkresën e mësipërme janë evidentuar orët për secilin anëtar të grupeve të punës, vlerë e
cila është brenda parashikimit të miratuar në buxhetin e vitit 2024 miratuar me vendim BA nr. 2 datë
23.02.2024.

Duke ju falenderuar për bashkëpunimini
ADMINISTRATOR
— Saimir

XHËLAJ

pe

Adresa /Adress: L. Pavarsia,
Bulevardi Vlorë-Skelë

Eeb-site: ësë Ymivlora.
Yalvlo cd. ë

e-mail: infocignivlora.cdutat
2

‘my?

Za.

a

REPUBLIKA

E

v

SHQIPËRISË

Va\t-

yor

UNIVERSITETI “ISMAIL QEMALI” VLORË
DREJTORIAE SIGURIMIT TË BRENDSHËM TË CILËSISË DHE PROJEKTEVE

Nr. 2Ç 8 p Prot.

Vlorë, më 9/12/2024

Lënda: Relacion mbi shkresën me nr.1978/2 prot, datë 11/12/2024

REKTORIT TË UNIVERSITETIT “ISMAIL QEMALI”, VLORË
PROF. DR AURELA SALIAJ

KËTU

Mbështetur në informacionin e përcjell me shkresë nr.2989 prot, datë 18.12.2024 si dhe nr. 1708
prot, datë 05.08.2024, të Dekanit të Fakultetit të Ekonomisë, për pagesën e anëtarve të personelit
akademik të përfshirë në project- propozimin me akronim WIKADIH, mbështetur në nenin 21
dhe 22 te Rregullores financiare të UV-së, miratuar me Vendim Bordi Administrimi Nr.45, date
05.12.2023, në vendimin e Senatit Akademik Nr.17, date 06.05.2021 “Per miratimin e
procedures se aplikimit, miratimit, implementimit dhe pageses e projekteve te financuara nga
programet e BE-se”, rezulton se grupi i punës ka zbatuar cdo detyrim që lidhet në kërkesën

pranë Rektorit për pagimin e orëve të punës prej BA-së për projekt propozimet e aplikuara në
vitin akademik 2023-2024, si më poshtë vijon:
Emri
Acronim

Emer mbiemer

Orët e punës së

|Donatorit ose | Viti

Kreditorit

Prof as Fioralba Vela

pagueshme

angazhimit
70

70

0

20

30

30

20

20

Viti
.

Prof As Zamira Sinaj

akademik

Dr. Rezarta Brokaj
Dr. Migena Petanaj

Oréte

aplikimit

23-24

Aplikuar
WIKADIH

E+ CBHE

shkurt 24

DREJTOR |

Rezarta sat

Adresa /Adress: L. Pavarésia,
Bulevardi Vloré-Skelé

Web-site: www.univlora.edu.al
Faqe | nga I

REPUBLIKA

E

SHQIPËRISË

UNIVERSITETI “ISMAIL QEMALI” VLORË
REKTORI

Vlorë më 39 (2% .2024
Lënda : Ngarkohen për zbatimin e shkresës bashkëlidhur në respektim të legjislacionit përkatës në fuqi :
Drejtuar:

> Administrator J
e

Drejtoria e Shërbimeve Mbështetëse

e

Sektori i Financës dhe Realizimit të Buxhetit

e

Sektori i Prokurimeve

e

Sektori i Shërbimeve

REKTORI

Prof. Dr. Aurela SALIAJ

Adresa /Adress: Bulevardi Vlore-Skele

Web-site: www.univlora.edu.al

Tel. Fax ++355222288

e-mail: info@univlora.edu.al
Faqe 1 nga I

The docs on the website are public, there is nothing confidential in them.

It is noticeable from the example that a simple OCR of the scanned PDF cannot not do a clean job at extracting the information. Especially the tabular data inside the document are broken.

To solve this, I tried to use specialized document processing tools, like Docling and Marker-PDF. These tools make an OCR first, and then utilize deep-learning models (such as Surya) in order to reconstruct the layout and structure of the original document.

5.1. Docling

Installation:
```
pipx install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pipx install docling[full]
```
I am installing the torch version that works only with CPU, because I don’t have any GPUs on my machine.

Test command:

docking --help
docling convert --help

docling convert \
    --ocr \
    --ocr-engine tesseract \
    --ocr-lang sqi \
    --ocr-mode full_page \
    --image-export-mode placeholder \
    vba7-2025.pdf

The output of the command:

2026-09-12 19:40:58,346 INFO    docling.document_converter: Going to convert document batch...
2026-09-12 19:40:58,347 INFO    docling.document_converter: Initializing pipeline for StandardPdfPipeline with opt
ions hash 3378e1b846087dde47fa6e0714d0ec48
2026-09-12 19:40:58,393 INFO    docling.models.utils.hf_model_download: Fetching model docling-project/docling-lay
out-heron (revision: main)...
2026-09-12 19:40:58,665 INFO    docling.models.utils.hf_model_download: Model docling-project/docling-layout-heron
 already cached at /root/.cache/huggingface/hub/models--docling-project--docling-layout-heron/snapshots/8f39ad3c0b
4c58e9c2d2c84a38465abf757272d8
Loading weights: 100%|███████████████████████████████████████████████████████| 770/770 [00:00<00:00, 6531.11it/s]
2026-09-12 19:40:59,436 INFO    docling.models.utils.hf_model_download: Fetching model docling-project/docling-mod
els (revision: v2.3.0)...
2026-09-12 19:40:59,591 INFO    docling.models.utils.hf_model_download: Model docling-project/docling-models alrea
dy cached at /root/.cache/huggingface/hub/models--docling-project--docling-models/snapshots/fc0f2d45e2218ea24bce50
45f58a389aed16dc23
2026-09-12 19:41:00,031 INFO    docling.pipeline.base_pipeline: Processing document vba7-2025.pdf

The result of the conversion:

<!-- image -->

<!-- image -->

<!-- image -->

## UNIVERSITETI "ISMAIL QEMALI" VLORË BORDI I ADMINISTRIMIT

Vlorë më,oj

<!-- image -->

<!-- image -->

## VENDIM

<!-- image -->

## PËR PAGESËN E ANËTARËVE TË GRUPEVE TË PUNËS PËR SHKRIM PROJEKTESH

Në mbështetje të nenit 49, pika 1, gërma ë, të Ligjit Nr. 8072015 "Për arsimin e lartë dhe kërkimin shkencor në institucionet e arsimit të lartë në Republikën e Shqipërisë", nenit 25, të Statutit të Universitetit "Ismail Qemali", Vlorë, Udhëzimit nr. 29, datë 10.09.2018 "Për veprimtarinë dhe ngarkesën mësimore të personelit akademik në institucionet e arsimit të lartë", si dhe referuar propozimit të Administratorit ardhur me shkresë 3122, prot, datë 26.12.2024, Bordi i Administrimit, në mbledhjen e datës 09.01.2025,

## VENDOSI :

1. Të miratojë pagesën e grupeve të punës sipas materialit bashkëlidhur dhe pjesë e këtij vendimi.
2. Ngarkohet për zbatimin e ketij vendimi Administratori dhe Drejtoria e Shërbimeve Mbështetëse të UVsë.

Ky vendim hyn në fuqi menjëherë.

KRYETAR

<!-- image -->

<!-- image -->

Nr. SI 3. L....Prot

<!-- image -->

## UNIVERSITETI "ISMAIL QEMALI" VLORË ADMINISTRATOR

Vlorë, më Vo, 11 2024

Lënda:

Përciellje për miratimin e efektit financiar për shkrimfhartim projektesh.

## BORDIT TË ADMINISTRIMIT TË UNIVERSITETIT "ISMAIL QEMALI'' VLORË Z. SOTIR NIKOLLA

Në zbatim të Ligjit nr. 8072015 "Për Arsimin e Lartë dhe Kërkimin Shkencor në Institucionet e Arsimit të Lartë në Republikën e Shqiperisë", Udhëzimit nr. 29 datë 10.09.2018 për "Veprimtarinë dhe ngarkesën mësimore të personelit akademik në IAL", ligjit nr. 9936 datë 26.06.2008 "Për menaxhimin e sistemit buxhetor në Republikën e Shqipërisë", i ndryshuar, ligjit nr. 10296 datë 08.07.2010 "Për menaxhimin financiar dhe kontrollin" ligjit nr. 97723 "Për buxhetin e vitit 2024", vendimit të Bordit të Administrimit nr. 8 datë 23.02.2024 "Për miratimin e buxhetit të Uinversitetit "Ismail Qemali" Vlorë për vitin 2024", Vendimin e Senatit Akademik nr. 17 datë 06.05.2021, vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 "Për miratimin e rregullores financiare të Universiteti "Ismail Qemali" Vlorë", si dhe shkresave të Drejtorisë së Sigurimit të Brendshëm të Cilësisë dhe Projekteve shkresa nr. 298971 prot., datë 20.12.2024 "Relacion mbi shkresën me nr. 197812 prot., datë 11.12.2024" efekti financiar për grupin e punës të cilat përmbushin çdo detyrim të vendimit tëSenatit Akademik nr. 17 datë 06.05.2021 dhe rregullores financiare miratuar me vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 është si më poshtë :

| Emer Mbiemer   | Pozicioni titulli” grada   | Projekti   | Vleralore   | Oret projekt )(   | Oret e punes se pagueshme   |   Vlera ) TotalejLeke |
|----------------|----------------------------|------------|-------------|-------------------|-----------------------------|-----------------------|
| Fjoralba Velaj | J Prof.As                  | VVIKADIH   | it          |                   |                             |                90,650 |
| Zamira Sinaj   | J Prof.As                  | (Aplikim   | 1,295       |                   | I I                         |                25,900 |
| Rezarta Brokaj | Dr                         | viti 2023- | 940 J)      |                   | I I                         |                28.200 |
| Migena Petanaj | Dr                         | 2024) J    | 940 )       |                   | I I                         |                18,800 |
| Totali         |                            |            |             |                   |                             |               163,550 |

Bazuar në shkresën e mësipërme janë evidentuar orët për secilin anëtar të grupeve të punës, vlerë e cila është brenda parashikimit të miratuar në buxhetin e vitit 2024 miratuar me vendim BA nr. 2 datë 23.02.2024.

Duke ju falenderuar për bashkëpunimini ADMINISTRATOR

<!-- image -->

<!-- image -->

<!-- image -->

## REPUBLIKA E SHQIPËRISË

## UNIVERSITETI "ISMAIL QEMALI" VLORË

## DREJTORIA E SIGURIMIT TË BRENDSHËM TË CILËSISË DHE PROJEKTEVE

<!-- image -->

Lënda:

Vlorë, më KI11212024

Relacion mbi shkresën me nr.197872 prot, datë 1171272024

<!-- image -->

## REKTORIT TË UNIVERSITETIT "ISMAIL QEMALI", VLORË

PROF. DR AURELA

## SALIAJ

KËTU

Mbështetur në informacionin e përcjell me shkresë nr.2989 prot, datë 18.12.2024 si dhe nr. 1708 prot, datë 05.08.2024, të Dekanit të Fakultetit të Ekonomisë, për pagesën e anëtarve të personelit akademik të përfshirë në projectpropozimin me akronim IVIKADIH, mbështetur në nenin 21 dhe 22 te Rregullores financiare të UV-së, miratuar me Vendim Bordi Administrimi Nr.45, date 05.12.2023, në vendimin e Senatit Akademik Nr.17, date 06.05.2021 "Per miratimin e procedures se aplikimit, miratimit, implementimit dhe pageses e projekteve te financuara nga programet e BE-se", rezulton se grupi i punës ka zbatuar cdo detyrim që lidhet në kërkesën pranë Rektorit për pagimin e orëve të punës prej BA-së për projekt propozimet e aplikuara në vitin akademik 2023-2024, si më poshtë vijon:

| Emer mbiemer          | Acronim   | Emri Donatorit ose Kreditorit   | Viti aplikimit   |   Orët ec angazhimit | Orët e punës së pagueshme,   |
|-----------------------|-----------|---------------------------------|------------------|----------------------|------------------------------|
| Prof as Fioralba Vela |           |                                 | Viti             |                   70 | 70,                          |
| Prof As Zamira Sinaj  |           |                                 | -                |                      |                              |
|                       |           |                                 | akademik         |                   20 | 20                           |
| Dr. Rezarta Brokaj    |           |                                 | 23-24            |                      |                              |
|                       |           |                                 | Aplikuar         |                   30 | 30,                          |
| Dr. Migena Petanaj    | VVIKADIH  | 1E4 CBHE                        | shkurt 24        |                      |                              |
|                       |           |                                 |                  |                   20 | 20                           |

DREJTOR

Rezarta Sinahali j

<!-- image -->

<!-- image -->

<!-- image -->

## REPUBLIKA E SHQIPËRISË

## UNIVERSITETI "ISMAIL QEMALI" VLORË REKTORI

REKTORI Prof. Dr. Aurela SALIAJ

<!-- image -->

<!-- image -->

Vlorë më 0 i .2024

Lënda : Ngarkohen për zbatimin e shkresës bashkëlidhur në respektim të legjislacionit përkatës në fuqi :

## Drejtuar:

- X Administrator
- -t Drejtoria e Shërbimeve Mbështetëse
- e Sektori i Financës dhe Realizimit të Buxhetit
- Sektori i Prokurimeve
- e Sektori i Shërbimeve

<!-- image -->

5.2. Marker-PDF

Installation:

pipx install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pipx install "marker-pdf[full]"

wget https://github.com/ggml-org/llama.cpp/releases/download/b10278/llama-b10278-bin-ubuntu-x64.tar.gz
tar xfz llama-b10278-bin-ubuntu-x64.tar.gz
cp llama-b10278/llama-server /usr/local/bin/
cp llama-b10278/*.so* /usr/local/bin/
llama-server --version

Testing:

marker_single --help | less

marker_single \
    vba7-2025.pdf \
    --output_dir ./ \
    --disable_image_extraction

The output of the command:

Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2026-09-12 20:42:56,664 [INFO] marker: Table processing stats: {'tables_ocr': 2, 'tables_total': 2}
2026-09-12 20:42:56,686 [INFO] marker: Saved markdown to ./vba7-2025
2026-09-12 20:42:56,687 [INFO] marker: Total time: 1035.676918745041
Force-killed llamacpp (pid 350394)

The result of the conversion:

REPUBLIKA E SHQIPËRISË

UNIVERSITETI "ISMAIL QEMALI" VLORË  
BORDI I ADMINISTRIMIT

Nr. 61/7 Prot.

Vlorë më, 09.01.2025

VENDIM

Nr. 2, datë 09.01.2025

**PËR PAGESËN E ANËTARËVE TË GRUPEVE TË PUNËS PËR SHKRIM  
PROJEKTESH**

Në mbështetje të nenit 49, pika 1, gërma ë, të Ligjit Nr. 80/2015 "Për arsimin e lartë dhe kërkimin shkencor në institucionet e arsimit të lartë në Republikën e Shqipërisë", nenit 25, të Statutit të Universitetit "Ismail Qemali", Vlorë, Udhëzimit nr. 29, datë 10.09.2018 "Për veprimtarinë dhe ngarkesën mësimore të personelit akademik në institucionet e arsimit të lartë", si dhe referuar propozimit të Administratorit ardhur me shkresë 3122, prot, datë 26.12.2024, Bordi i Administrimit, në mbledhjen e datës 09.01.2025,

**VENDOSI :**

1. 1. Të miratojë pagesën e grupeve të punës sipas materialit bashkëlidhur dhe pjesë e këtij vendimi.
2. 2. Ngarkohet për zbatimin e ketij vendimi Administratori dhe Drejtoria e Shërbimeve Mbështetëse të UV- së.

Ky vendim hyn në fuqi menjëherë.

**K R Y E T A R**

**SOTIR NIKOLLA**

REPUBLIKA E SHQIPËRISË

**UNIVERSITETI "ISMAIL QEMALI" VLORË  
ADMINISTRATOR**

Nr.. 31.2.2... Prot

Vlorë, më 26. 11. 2024

**Lënda:** Përciellje për miratimin e efektit financiar për shkrim/hartim projektesh.

**BORDIT TË ADMINISTRIMIT TË UNIVERSITETIT "ISMAIL QEMALI" VLORË  
Z. SOTIR NIKOLLA**

Në zbatim të Ligjit nr. 80/2015 "Për Arsimin e Lartë dhe Kërkimin Shkencor në Institucionet e Arsimit të Lartë në Republikën e Shqiperisë", Udhëzimit nr. 29 datë 10.09.2018 për "Veprimtarinë dhe ngarkesën mësimore të personelit akademik në IAL", ligjit nr. 9936 datë 26.06.2008 "Për menaxhimin e sistemit buxhetor në Republikën e Shqipërisë", i ndryshuar, ligjit nr. 10296 datë 08.07.2010 "Për menaxhimin financiar dhe kontrollin" ligjit nr. 97/23 "Për buxhetin e vitit 2024", vendimit të Bordit të Administrimit nr. 8 datë 23.02.2024 "Për miratimin e buxhetit të Universitetit "Ismail Qemali" Vlorë për vitin 2024", Vendimin e Senatit Akademik nr. 17 datë 06.05.2021, vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 "Për miratimin e rregullores financiare të Universitetit "Ismail Qemali" Vlorë", si dhe shkresave të Drejtorisë së Sigurimit të Brendshëm të Cilësisë dhe Projekteve shkresa nr. 2989/1 prot., datë 20.12.2024 "Relacion mbi shkresën me nr. 1978/2 prot., datë 11.12.2024" efekti financiar për grupin e punës të cilat përmbushin çdo detyrim të vendimit të Senatit Akademik nr. 17 datë 06.05.2021 dhe rregullores financiare miratuar me vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 është si më poshtë :

| Nr | Emer Mbiemer   | Pozicioni/<br>titulli/<br>grada | Projekti            | Vlera/ore | Oret projekt | Oret e punes<br>se pagueshme | Vlera<br>Totale/Leke |
|----|----------------|---------------------------------|---------------------|-----------|--------------|------------------------------|----------------------|
| 1  | Fjoralba Velaj | Prof.As                         | WIKADIH             | 1,295     | 70           | 70                           | 90,650               |
| 2  | Zamira Sinaj   | Prof.As                         | (Aplikim            | 1,295     | 20           | 20                           | 25,900               |
| 3  | Rezarta Brokaj | Dr                              | viti 2023-<br>2024) | 940       | 30           | 30                           | 28,200               |
| 4  | Migena Petanaj | Dr                              |                     | 940       | 20           | 20                           | 18,800               |
| 4  | <b>Totali</b>  |                                 |                     |           | <b>140</b>   | <b>140</b>                   | <b>163,550</b>       |

Bazuar në shkresën e mësipërme janë evidentuar orët për secilin anëtar të grupeve të punës, vlerë e cila është brenda parashikimit të miratuar në buxhetin e vitit 2024 miratuar me vendim BA nr. 2 datë 23.02.2024.

Duke ju falenderuar për bashkëpunimin!

ADMINISTRATOR

REPUBLIKA E SHQIPËRISË  
 UNIVERSITETI "ISMAIL QEMALI" VLORË  
 DREJTORIA E SIGURIMIT TË BRENDSHËM TË CILËSISË DHE PROJEKTEVE

Projekt  
 23.12.2024  
 [Signature]

Nr. 298P / Prot.  
 /

Vlorë, më 27/12/2024

Lënda: Relacion mbi shkresën me nr.1978/2 prot, datë 11/12/2024

**REKTORIT TË UNIVERSITETIT "ISMAIL QEMALI", VLORË**

**PROF. DR AURELA SALIAJ**

**KËTU**

Mbështetur në informacionin e përcjell me shkresë nr.2989 prot, datë 18.12.2024 si dhe nr. 1708 prot, datë 05.08.2024, të Dekanit të Fakultetit të Ekonomisë, për pagesën e anëtarve të personelit akademik të përfshirë në projekt-propozimin me akronim WIKADIH, mbështetur në nenin 21 dhe 22 te Rregullores financiare të UV-së, miratuar me Vendim Bordi Administrimi Nr.45, date 05.12.2023, në vendimin e Senatit Akademik Nr.17, date 06.05.2021 "Per miratimin e procedures se aplikimit, miratimit, implementimit dhe pageses e projekteve te financuara nga programet e BE-se", rezulton se grupi i punës ka zbatuar cdo detyrim që lidhet në kërkesën pranë Rektorit për pagimin e orëve të punës prej BA-së për projekt propozimet e aplikuara në vitin akademik 2023-2024, si më poshtë vijon:

| Emer mbiemer          | Acronim | Emri Donatorit ose Kreditorit | Viti aplikimit                                     | Orët e angazhimit | Orët e punës së pagueshme |    |    |
|-----------------------|---------|-------------------------------|----------------------------------------------------|-------------------|---------------------------|----|----|
| Prof as Fioralba Vela | WIKADIH | E+ CBHE                       | Viti<br>akademik<br>23-24<br>Aplikuar<br>shkurt 24 | 70                | 70                        |    |    |
| Prof As Zamira Sinaj  |         |                               |                                                    |                   |                           |    |    |
| Dr. Rezarta Brokaj    |         |                               |                                                    |                   |                           | 20 | 20 |
| Dr. Migena Petanaj    |         |                               |                                                    |                   |                           | 30 | 30 |
|                       |         |                               |                                                    |                   |                           | 20 | 20 |

**DREJTOR**

Rezarta Sinahaliaj

![]()REPUBLIKA E SHQIPËRISË  
UNIVERSITETI "ISMAIL QEMALI" VLORË  
REKTORI

Vlorë më, 20 | 2 . 2024

**Lënda :** Ngarkohen për zbatimin e shkresës bashkëlidhur në respektim të legjislacionit përkatës në fuqi :

**Drejtuar:**

➤ Administrator ✓

- • Drejtoria e Shërbimeve Mbështetëse
- • Sektori i Financës dhe Realizimit të Buxhetit
- • Sektori i Prokurimeve
- • Sektori i Shërbimeve

**REKTORI**  
**Prof. Dr. Aurela SALIAJ**

Note that it is taking aout 17 min just for processing one document (Total time: 1035.676918745041).

Both Docling and Marker-PDF download open-source models from Hugging Face and run them locally using PyTorch.

Testing with a Small Local Model:

# install and test ollama
apt install -y curl ca-certificates
curl -fsSL https://ollama.com/install.sh | sh
systemctl status ollama
ollama pull qwen2.5vl
ollama list
ollama run qwen2.5vl

marker_single \
    vba7-2025.pdf \
    --output_dir ./ \
    --disable_image_extraction\
    --use_llm \
    --llm_service marker.services.ollama.OllamaService \
    --ollama_base_url http://localhost:11434 \
    --ollama_model qwen2.5vl

The output of the command:

Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2026-09-13 15:47:37,270 [INFO] marker: Table processing stats: {'tables_ocr': 2, 'tables_total': 2}
LLMTableProcessor running:  50%|█████████████████████████                         | 1/2 [02:58<02:58, 178.93s/it]2026-09-13 15:53:21,496 [INFO] marker: Table parsing warning: too many columns found
2026-09-13 15:53:21,496 [INFO] marker: Table parsing warning: too many columns found
2026-09-13 15:53:21,496 [INFO] marker: Table parsing warning: too many columns found
2026-09-13 15:53:21,497 [INFO] marker: Table parsing warning: too many columns found
LLMTableProcessor running: 100%|██████████████████████████████████████████████████| 2/2 [05:44<00:00, 172.11s/it]
LLM processors running: 100%|████████████████████████████████████████████████████| 13/13 [09:58<00:00, 46.07s/it]
Running LLMSectionHeaderProcessor: 100%|███████████████████████████████████████████| 1/1 [00:39<00:00, 39.48s/it]
2026-09-13 16:03:59,913 [INFO] marker: Saved markdown to ./vba7-2025
2026-09-13 16:03:59,913 [INFO] marker: Total time: 2011.3281693458557
Force-killed llamacpp (pid 416339)

The result of the conversion:

Image /page/0/Picture/19 description: The image shows a circular emblem with the letters 'UV' prominently displayed in the center. Surrounding the 'UV' are the words 'Virtus', 'Scientia', and 'Veritas', arranged in a circular fashion. Below the 'UV' and the surrounding words, the text reads 'Universitati 

Image /page/0/Picture/20 description: The image depicts a shield-shaped emblem with a black and white design. At the center of the shield is a stylized black eagle with outstretched wings, facing to the left. The background of the shield is divided into two sections: the upper section is white, and the lower section is black. The eagle appears to be in flight, with its talons extended, and its head turned slightly to the side. The overall design is symmetrical and has a heraldic appearance.

REPUBLIKA E SHQIPËRISË

UNIVERSITETI "ISMAIL QEMALI" VLORË  
BORDI I ADMINISTRIMIT

Nr. 61/7 Prot.

Vlorë më, 09.01.2025

VENDIM

Nr. 2, datë 09.01.2025

**PËR PAGESËN E ANËTARËVE TË GRUPEVE TË PUNËS PËR SHKRIM  
PROJEKTESH**

Në mbështetje të nenit 49, pika 1, gërma ë, të Ligjit Nr. 80/2015 "Për arsimin e lartë dhe kërkimin shkencor në institucionet e arsimit të lartë në Republikën e Shqipërisë", nenit 25, të Statutit të Universitetit "Ismail Qemali", Vlorë, Udhëzimit nr. 29, datë 10.09.2018 "Për veprimtarinë dhe ngarkesën mësimore të personelit akademik në institucionet e arsimit të lartë", si dhe referuar propozimit të Administratorit ardhur me shkresë 3122, prot, datë 26.12.2024, Bordi i Administrimit, në mbledhjen e datës 09.01.2025,

**VENDOSI :**

1. 1. Të miratojë pagesën e grupeve të punës sipas materialit bashkëlidhur dhe pjesë e këtij vendimi.
2. 2. Ngarkohet për zbatimin e ketij vendimi Administratori dhe Drejtoria e Shërbimeve Mbështetëse të UV- së.

Ky vendim hyn në fuqi menjëherë.

**K R Y E T A R**

**SOTIR NIKOLLA**

Image /page/0/Picture/34 description: The image shows a document with a signature and a seal. The name 'SOTIR NIKOLLA' is printed at the top. Below the name, there is a blue signature. To the right, there is a circular seal with the text 'REPUBLIKA E SHQIPERISE' (Republic of Albania) and 'MINISTRIA E PUNES' (Ministry of Labor). The document appears to be official, possibly a certificate or an official document from Albania.

Image /page/1/Picture/17 description: The image shows a circular emblem with the text 'Virtus - Scientia - Veritas' arranged in a circular fashion around the perimeter. In the center, there is a stylized 'UV' with a swoosh underneath it. Below the swoosh, the text 'universiteti 

Image /page/1/Picture/18 description: The image depicts a shield-shaped emblem featuring a black double-headed eagle with outstretched wings. The eagle is centered within the shield, which has a textured, possibly fabric-like background. The design is simple and symmetrical, with the eagle's head facing forward and its wings spread wide, symbolizing strength and vigilance.

REPUBLIKA E SHQIPËRISË

**UNIVERSITETI "ISMAIL QEMALI" VLORË  
ADMINISTRATOR**

Nr.. 31.2.2... Prot

Vlorë, më 26. 11. 2024

**Lënda:** Përciellje për miratimin e efektit financiar për shkrim/hartim projektesh.

**BORDIT TË ADMINISTRIMIT TË UNIVERSITETIT "ISMAIL QEMALI" VLORË  
Z. SOTIR NIKOLLA**

Në zbatim të Ligjit nr. 80/2015 "Për Arsimin e Lartë dhe Kërkimin Shkencor në Institucionet e Arsimit të Lartë në Republikën e Shqiperisë", Udhëzimit nr. 29 datë 10.09.2018 për "Veprimtarinë dhe ngarkesën mësimore të personelit akademik në IAL", ligjit nr. 9936 datë 26.06.2008 "Për menaxhimin e sistemit buxhetor në Republikën e Shqipërisë", i ndryshuar, ligjit nr. 10296 datë 08.07.2010 "Për menaxhimin financiar dhe kontrollin" ligjit nr. 97/23 "Për buxhetin e vitit 2024", vendimit të Bordit të Administrimit nr. 8 datë 23.02.2024 "Për miratimin e buxhetit të Universitetit "Ismail Qemali" Vlorë për vitin 2024", Vendimin e Senatit Akademik nr. 17 datë 06.05.2021, vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 "Për miratimin e rregullores financiare të Universitetit "Ismail Qemali" Vlorë", si dhe shkresave të Drejtorisë së Sigurimit të Brendshëm të Cilësisë dhe Projekteve shkresa nr. 2989/1 prot., datë 20.12.2024 "Relacion mbi shkresën me nr. 1978/2 prot., datë 11.12.2024" efekti financiar për grupin e punës të cilat përmbushin çdo detyrim të vendimit të Senatit Akademik nr. 17 datë 06.05.2021 dhe rregullores financiare miratuar me vendimit të Bordit të Administrimit nr. 45 datë 05.12.2023 është si më poshtë :

| Nr | Emer Mbiemer   | Pozicioni/<br>titulli/<br>grada | Projekti            | Vlera/ore | Oret projekt | Oret e punes<br>se pagueshme | Vlera<br>Totale/Leke |
|----|----------------|---------------------------------|---------------------|-----------|--------------|------------------------------|----------------------|
| 1  | Fjoralba Velaj | Prof.As                         | WIKADIH             | 1,295     | 70           | 70                           | 90,650               |
| 2  | Zamira Sinaj   | Prof.As                         | (Aplikim            | 1,295     | 20           | 20                           | 25,900               |
| 3  | Rezarta Brokaj | Dr                              | viti 2023-<br>2024) | 940       | 30           | 30                           | 28,200               |
| 4  | Migena Petanaj | Dr                              |                     | 940       | 20           | 20                           | 18,800               |
| 4  | Totali         |                                 |                     |           | 140          | 140                          | 163,550              |

Bazuar në shkresën e mësipërme janë evidentuar orët për secilin anëtar të grupeve të punës, vlerë e cila është brenda parashikimit të miratuar në buxhetin e vitit 2024 miratuar me vendim BA nr. 2 datë 23.02.2024.

Duke ju falenderuar për bashkëpunimin!

ADMINISTRATOR

Image /page/1/Picture/33 description: The image appears to be a close-up of a document or certificate with text and a signature. The text includes names, email addresses, and a logo or seal. The visible text includes 'Saimir XHELA', 'univlora.edu.al', and 'info@univlora.edu.al'. There is also a circular seal or stamp with text around it, but the text is partially obscured and difficult to read. The document seems to be related to an administrative or educational institution.

Image /page/2/Picture/16 description: The image shows a circular emblem with the text 'Universidad de las Américas' written around the outer edge. The center of the emblem features the letters 'UV' in a stylized font, with a wavy line underneath. The design is monochromatic and appears to be a seal or logo.

Image /page/2/Picture/17 description: The image appears to be a close-up of a textured surface with a pattern that resembles a shield or emblem. The design includes a central motif that looks like a stylized tree or plant with branches extending outward. The background has a ribbed or segmented pattern, possibly resembling a rib cage or a series of parallel lines. The overall color scheme is monochromatic, with shades of gray and black.

REPUBLIKA E SHQIPËRISË  
 UNIVERSITETI "ISMAIL QEMALI" VLORË  
 DREJTORIA E SIGURIMIT TË BRENDSHËM TË CILËSISË DHE PROJEKTEVE

Projekt  
 23.12.2024  
 [Signature]

Nr. 298P / Prot.  
 /

Vlorë, më 27/12/2024

Lënda: Relacion mbi shkresën me nr.1978/2 prot, datë 11/12/2024

**REKTORIT TË UNIVERSITETIT "ISMAIL QEMALI", VLORË**

**PROF. DR AURELA SALIAJ**

**KËTU**

Mbështetur në informacionin e përcjell me shkresë nr.2989 prot, datë 18.12.2024 si dhe nr. 1708 prot, datë 05.08.2024, të Dekanit të Fakultetit të Ekonomisë, për pagesën e anëtarve të personelit akademik të përfshirë në projekt-propozimin me akronim WIKADIH, mbështetur në nenin 21 dhe 22 te Rregullores financiare të UV-së, miratuar me Vendim Bordi Administrimi Nr.45, date 05.12.2023, në vendimin e Senatit Akademik Nr.17, date 06.05.2021 "Per miratimin e procedures se aplikimit, miratimit, implementimit dhe pageses e projekteve te financuara nga programet e BE-se", rezulton se grupi i punës ka zbatuar cdo detyrim që lidhet në kërkesën pranë Rektorit për pagimin e orëve të punës prej BA-së për projekt propozimet e aplikuara në vitin akademik 2023-2024, si më poshtë vijon:

| Emer mbiemer          | Acronim | Emri Donatorit ose Kreditorit | Viti aplikimit                                     | Orët e angazhimit | Orët e punës së pagueshme |    |    |
|-----------------------|---------|-------------------------------|----------------------------------------------------|-------------------|---------------------------|----|----|
| Prof as Fioralba Vela | WIKADIH | E+ CBHE                       | Viti<br>akademik<br>23-24<br>Aplikuar<br>shkurt 24 | 70                | 70                        |    |    |
| Prof As Zamira Sinaj  |         |                               |                                                    |                   |                           |    |    |
| Dr. Rezarta Brokaj    |         |                               |                                                    |                   |                           | 20 | 20 |
| Dr. Migena Petanaj    |         |                               |                                                    |                   |                           | 30 | 30 |
|                       |         |                               |                                                    |                   |                           | 20 | 20 |

**DREJTOR**

Rezarta Sinahaliaj

# BRETOR

**Dita Sinahaliaj**

Image /page/3/Picture/15 description: The image is a black and white illustration featuring a stylized figure of a person holding a flag. The figure is depicted in a dynamic pose, with one arm raised and the other bent at the elbow, holding a flag with a visible emblem. The number '80' is prominently displayed in a large, stylized font, with the word 'VJETORI' written along the curve of the number. The overall design has a bold and graphic quality.

Image /page/3/Picture/16 description: The image depicts a shield emblem featuring a double-headed eagle. The eagle is detailed with intricate feather patterns and is positioned centrally within the shield. The shield itself has a textured border, giving it a medieval or heraldic appearance.

REPUBLIKA E SHQIPËRISË  
UNIVERSITETI "ISMAIL QEMALI" VLORË  
REKTORI

Image /page/3/Picture/18 description: The image features a circular emblem with the letters 'UV' prominently displayed in the center. Surrounding the 'UV' are the words 'Virtus', 'Scientia', and 'Veritas', arranged in a circular fashion. Below the 'UV' and within the circle, the text 'Ismail Ogemu' is inscribed. The design appears to be a logo or seal, possibly representing an institution or organization.

Vlorë më, 20 | 2 . 2024

**Lënda :** Ngarkohen për zbatimin e shkresës bashkëlidhur në respektim të legjislacionit përkatës në fuqi :

**Drejtuar:**

➤ Administrator ✓

- • Drejtoria e Shërbimeve Mbështetëse
- • Sektori i Financës dhe Realizimit të Buxhetit
- • Sektori i Prokurimeve
- • Sektori i Shërbimeve

**REKTORI**  
**Prof. Dr. Aurela SALIAJ**

Image /page/3/Picture/25 description: The image shows a circular stamp with text in Albanian. The text reads 'STAMP OF THE REPUBLIKAE SHQIPERIE - UNIVERSITY OF VLORE' and includes a date '2022'. The design features a circular emblem in the center with a star-like pattern.

This takes more than 30 min (Total time: 2011.3281693458557) and it is analysing and describing in text the various images that appear in the document.

Running these heavy transformer models on my CPU-only server without an NVIDIA GPU leeds to a massive bottleneck: a single, complex, scanned PDF could take more than 15-30 minutes to parse and convert into clean Markdown.

6. Running Local LLMs

I mentioned previously that self-hosted platforms, like AnythingLLM, Open WebUI, RAGFlow, PipesHub, etc. do not provide AI models. They act as the orchestration and database layer, but we must supply:

An Embedding Model to vectorize the documents.
An LLM Model to synthesize the retrieved context and formulate the response.

These models can be provided:

By purchasing a service (tokens) from providers like Google, OpenAI, Anthropic, and many others. Usually we provide the URL of the service, the API key, the name of the model, etc.
Running open-weights models locally, usually through Ollama.

I ignored the first option, because I don’t have any budged for this project, and focused my efforts on the second option: running open-weights models locally.

We already saw how to install Ollama (on the previous section):

# install and test ollama
apt install -y curl ca-certificates
curl -fsSL https://ollama.com/install.sh | sh

systemctl status ollama

ollama pull llama3.2:1b
ollama pull gemma3:1b
ollama list

ollama run gemma3:1b

We can find available models at:

Ollama is running as a service, and it can be accessed from other local applications (AnythingLLM etc.) at the url http://localhost:11434

If the application that needs to access a model through ollama is installed on another host, we can make ollama available through a reverse proxy.

I use nginx as a reverse proxy and my configuration looks like this:

map $http_authorization $valid_api_key {
    default 0;
    "Bearer key1-v7gahShaiBise" 1;
    "Bearer key2-sis7Hheiphei2" 1;
    "Bearer key3-eeVeem7lau9Oo" 1;
}

server {
    listen 443 ssl;
    server_name ollama.univlora.al ;

    ssl_certificate     certs/ollama.univlora.al/fullchain.pem;
    ssl_certificate_key certs/ollama.univlora.al/privkey.pem;

    access_log /var/log/nginx/ollama.univlora.al-access.log;
    error_log /var/log/nginx/ollama.univlora.al-error.log error;

    client_max_body_size 100M;

    location / {
        # Enforce API Key verification
        if ($valid_api_key = 0) {
            return 401 '{"error": "Unauthorized: Invalid or missing API Key"}';
        }

        proxy_pass http://localhost:11434;

        # Required for LLM streaming & host header validation
        proxy_http_version 1.1;
        proxy_set_header Host localhost;
        proxy_set_header Connection "";
        proxy_buffering off;

        # Prevent timeouts during long generation runs
        proxy_read_timeout 600s;
        proxy_connect_timeout 600s;
        proxy_send_timeout 600s;
    }
}

If the reverse proxy is not on the same host as ollama, then http://localhost:11434; should be http://10.120.182.15:11434; (where 10.120.182.15 is the IP of the ollama host).

In this case we should also make sure that the port 11434 on the ollama host is open, and the ollama service is listening to any interface (by default it listens only to the local interface):
1. Run: systemctl edit ollama
2. Add these lines to the config file:
  [Service] Environment="OLLAMA_HOST=0.0.0.0" Environment="OLLAMA_ORIGINS=*"
3. Restart the service:
  systemctl daemon-reload systemctl restart ollama

To test it, we can use a curl command like this:

curl https://ollama.univlora.al/v1/chat/completions \
  -H "Authorization: Bearer key1-v7gahShaiBise" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma3:1b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

7. Which Model to Use?

The machine that I can use for running local models is only an old Fujitsu PRIMERGY server without dedicated GPUs, so it has to rely entirely on CPU-based inference. Tweaking the settings to run local models as efficiently as possible is an art on its own, depending on NUMA topology of the server, CPUs, RAM allocations, etc.

For example the command numactl --hardware shows something like this:

available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13
node 0 size: 55371 MB
node 0 free: 53764 MB
node 1 cpus: 14 15 16 17 18 19 20 21 22 23 24 25 26 27
node 1 size: 55396 MB
node 1 free: 53873 MB
node distances:
node     0    1
   0:   10   20
   1:   20   10

Because CPU-based LLM inference is strictly memory-bandwidth bound (rather than CPU-compute bound), the RAM bandwidth is the single bottleneck that dictates the speed of token generation.

To measure the performance of the server we can use the benchmarking tool STREAM:

wget -q https://www.cs.virginia.edu/stream/FTP/Code/stream.c
gcc -O3 -fopenmp -mcmodel=large -DSTREAM_ARRAY_SIZE=100000000 -DNTIMES=20 stream.c -o stream
chmod +x stream

export OMP_NUM_THREADS=28
export OMP_PROC_BIND=spread
./stream

The output of ./stream:

Function    Best Rate MB/s  Avg time     Min time     Max time
Copy:           45352.7     0.035442     0.035279     0.036387
Scale:          35568.3     0.045251     0.044984     0.046060
Add:            40922.0     0.059030     0.058648     0.059744
Triad:          40847.5     0.059010     0.058755     0.059488

Gemini makes a calculation/prediction like this:

The benchmark result (~45 GB/s memory bandwidth) gives a clear mathematical picture of how this machine will perform:

Model Size	RAM Footprint	Estimated Speed	Usability
8B	~5.0 GB	~8.5 tokens/sec	Smooth
14B	~9.5 GB	~4.5 tokens/sec	Comfortable
32B	~20.0 GB	~2.1 tokens/sec	Sluggish
72B	~47.0 GB	~0.9 tokens/sec	Very slow

Model Size

RAM Footprint

Estimated Speed

Usability

~5.0 GB

~8.5 tokens/sec

Smooth

14B

~9.5 GB

~4.5 tokens/sec

Comfortable

32B

~20.0 GB

~2.1 tokens/sec

Sluggish

72B

~47.0 GB

~0.9 tokens/sec

Very slow

Because the machine has 28 vCPUs split across 2 NUMA nodes, Ollama needs to be tuned so it doesn’t suffer from cross-socket memory bottlenecks:

Edit its configuration:
```
systemctl edit ollama
```

Add this line under [Service]:

[Service]
Environment="OLLAMA_NUM_THREADS=28"

Restart:

systemctl daemon-reload
systemctl restart ollama

For large models (32B–70B) that need all of RAM, run Ollama normally so it spans both NUMA nodes:
```
ollama run qwen2.5:32b
```
For smaller models, crossing the UPI bus between the two sockets adds unnecessary latency. Running Ollama pinned to a single NUMA node yields 15–20% faster token generation for 8B models:
```
# Force Ollama CLI to use only NUMA Node 0 (CPUs 0-13 and local Node 0 RAM)
numactl --cpunodebind=0 --membind=0 ollama run llama3.1:8b
```

The models that Gemini recommends

Based on my server and its performance, Gemini makes recommendations like this:

Tier 1: "Everyday Fast" (7–10 tokens/sec)

Ideal for quick Q&A, rapid chat, daily assistance, and lightweight summaries.

llama3.1:8b (Meta): The industry standard small model. Excellent instruction-following, very responsive on 14 vCPUs.
qwen2.5:7b (Alibaba): Exceptional multilingual capabilities, strong structured JSON formatting, and solid general knowledge.
deepseek-r1:8b (DeepSeek): A reasoning model fine-tuned for step-by-step logic, math, and code debugging. Uses chain-of-thought <think> tags before answering.

Tier 2: The "Sweet Spot" (4–5 tokens/sec)

The best balance of reasoning depth, coding capabilities, and practical CPU generation speed.

qwen2.5:14b or qwen2.5-coder:14b: The single best model class for CPU execution. Outperforms many older 32B/70B models in coding, logic, and long-context processing while generating output at a comfortable reading pace.
deepseek-r1:14b: Combines deep reasoning with manageable size. Solves complex logic puzzles and code architectural problems far better than 8B models.
phi4:14b (Microsoft): A heavily optimized model that punches well above its weight class in STEM, reasoning, and technical comprehension.

8. Broken Albanian Language

Our chatbot (RAG system) must understand questions and formulate highly precise answers in the Albanian language. This requirement immediately eliminates many of the most famous models.

This is what Gemini tells me (roughly):

Meta’s Llama 3.1 is officially optimized for only eight major languages. When forced to answer in Albanian, it frequently makes severe grammatical errors, uses awkward literal translations from English, or randomly slips into Italian or English during complex synthesis.

Furthermore, Llama’s tokenizer is highly inefficient for Albanian text. Because languages like Albanian contain specific characters (ë, ç) and distinct agglutinative structures, Llama’s tokenizer splits these words into many tiny fragments (sub-tokens) [88].

Alibaba’s Qwen 2.5 family is supposed to be better, trained on 18 trillion tokens with a custom 151,000-token multilingual vocabulary. For local CPU execution, Qwen 2.5 14B should be the "sweet spot," delivering exceptional reasoning and native language fluency at a comfortable reading speed of roughly 4.5 tokens per second.

From the tests that I did with feasible models (those that are able to run on my server, even slowly), I found out that none of them has sufficient skills in Albanian (especially when composing an answer).

This is where the project hits a wall, because no matter how well the other parts of the pipeline work (crawling the website, PDF procesing, etc.), the system will never be able to formulate a decent answer in Albanian.

So, the resources that are currently available, are not enough to completely self-host a project like this.

9. The True Cost of Self-Hosting

I wonder, if (hypothetically) I had no money restrictions, what would be the best open-weight models that perform well in Albanian? How much resources would they need to run properly? And how much would cost the hardware for hosting them?

I asked Gemini:

If I had a good enough server (enough resources, RAM, computational power, GPU, fast, etc.) which models would you suggest, that perform good in Albanian, for example as good as Gemini? How much resources, and what kind of server would be required by the best models?

If you remove resource constraints and step into enterprise-grade GPU server territory, you no longer need to rely on CPU-bound execution or 8B–14B models. You can host open-weight models that match or exceed the performance of proprietary models like Gemini 1.5/2.0, GPT-4o, or Claude 3.5 Sonnet in non-English languages like Albanian.

1. Top Open-Weight Models for Albanian (Gemini-Level Performance)

1. DeepSeek-V3 / DeepSeek-R1 (671B Mixture-of-Experts)

Why it matches Gemini: DeepSeek-V3 (and its reasoning counterpart, R1) currently stands at the top of open-weight LLMs globally. It matches GPT-4o and Gemini 1.5 Pro in logic, coding, and multilingual understanding.
Albanian Performance: Exceptional. Its massive knowledge base and advanced Mixture-of-Experts (MoE) architecture allow it to process Albanian nuances, complex legal texts, and administrative documents with fluid, near-native grammar.

2. Qwen 2.5 72B (72B Dense)

Why it matches Gemini: Alibaba’s Qwen family is widely regarded as the strongest non-English foundation model series. Pre-trained on over 18 trillion tokens with a custom 151,000-token multilingual vocabulary, it avoids word-fragmentation issues in Albanian.
Albanian Performance: Extremely fluent. For document summarization, extraction, and RAG in Albanian, Qwen 2.5 72B is arguably the cleanest and most reliable dense model available.

3. Llama 3.1 405B / Llama 3.3 70B (Meta)

Why it matches Gemini: Llama 3.1 405B was Meta’s first ultra-large open model built to directly compete with GPT-4. Llama 3.3 70B offers similar intelligence at a fraction of the parameter count.
Albanian Performance: Strong reasoning and synthesis, though Qwen and DeepSeek have a slight edge in native Albanian vocabulary choices due to their tokenizer architectures.

2. Server & Hardware Requirements

To run these top-tier models at full speed (50 to 100+ tokens/second), the bottleneck shifts entirely to GPU VRAM capacity and inter-GPU interconnect bandwidth.

Tier 1: The Frontier Server (To run DeepSeek-V3 / DeepSeek-R1)

To run a 671B MoE model natively in FP8 precision, you need a dedicated multi-GPU AI server cluster.

  ┌─────────────────────────────────────────────────────────────┐
  │                 ENTERPRISE GPU NODE (8x GPU)                │
  │  ┌───────┐ ┌───────┐ ┌───────┐ ┌───────┐ ... ┌───────┐      │
  │  │ GPU 0 │ │ GPU 1 │ │ GPU 2 │ │ GPU 3 │     │ GPU 7 │      │
  │  │(141GB)│ │(141GB)│ │(141GB)│ │(141GB)│     │(141GB)│      │
  │  └───┬───┘ └───┬───┘ └───┬───┘ └───┬───┘     └───┬───┘      │
  │      └─────────┴──── NVLink 4 ─────┴─────────────┘          │
  └─────────────────────────────────────────────────────────────┘

GPUs: 8x NVIDIA H200 (141GB VRAM each = ~1,128 GB VRAM) OR 8x AMD Instinct MI300X (192GB VRAM each).
GPU Interconnect: NVLink 4 / NVSwitch (900 GB/s to 1.8 TB/s inter-GPU bandwidth). High-speed interconnect is critical so the 8 GPUs can swap MoE expert routing tokens instantly.
System RAM: 1 TB to 2 TB DDR5 ECC RAM.
CPUs: Dual AMD EPYC 9004/9005 Series (e.g., 2x 64-core or 128-core CPUs).
Storage: 4TB Enterprise PCIe Gen5 NVMe SSD (U.2/U.3) for fast model loading (weights exceed 700 GB).
Estimated Hardware Cost: ~$200,000 – $350,000+ (or ~$15–$25/hour on cloud providers like Lambda Labs, RunPod, or Azure).

Tier 2: The Enterprise Workhorse (To run Qwen 2.5 72B / Llama 3.3 70B in Unquantized FP16)

To run 70B-class models with zero quantization quality loss at maximum throughput.

GPUs: 2x to 4x NVIDIA A100 (80GB) / H100 (80GB) OR 2x NVIDIA RTX 6000 Ada / Blackwell (48GB–96GB). (Requires ~160 GB total VRAM for unquantized FP16, or ~80 GB for FP8).
GPU Interconnect: NVLink bridge between GPUs.
System RAM: 256 GB to 512 GB DDR5 RAM.
CPUs: Dual Intel Xeon Scalable or AMD EPYC (32 to 64 physical cores).
Storage: 2TB NVMe PCIe Gen4/Gen5 SSD.
Estimated Hardware Cost: ~$25,000 – $60,000.

Tier 3: The Mid-Tier GPU Server (To run Qwen 2.5 72B in 4-bit/8-bit Quantized)

If you want to run 72B models on a single workstation or 4U rack server without spending enterprise-cloud budgets.

GPUs: 4x NVIDIA RTX 4090 (24GB VRAM = 96GB total) OR 2x NVIDIA RTX 6000 Ada (48GB VRAM = 96GB total).
System RAM: 128 GB to 256 GB DDR5 RAM.
CPUs: Single AMD Threadripper or high-end Intel Xeon.
Estimated Hardware Cost: ~$8,000 – $18,000.

3. How the Software Stack Changes on GPU Hardware

When running on dedicated GPU hardware, you replace Ollama/CPU drivers with high-throughput enterprise inference engines:

vLLM / TensorRT-LLM / SGLang: Instead of processing one request at a time, these engines use PagedAttention and Continuous Batching. A single server with 8x H200 GPUs running vLLM can serve hundreds of concurrent users typing in Albanian simultaneously without latency drops.
Tensor Parallelism (TP) & Expert Parallelism (EP): The model’s neural network layer matrices are split dynamically across all 8 GPUs. Each GPU computes its fraction of the token math in parallel, delivering responses at 80+ tokens per second.

10. Conclusions and Thoughts

Building a self-hosted RAG system for the university involves balancing several trade-offs. Given that all documentation and queries must be in Albanian, fully self-hosting open-weight models that write fluently in Albanian requires expensive hardware (GPUs and RAM) that is out of reach. Switching to cloud LLM providers (e.g., OpenAI or Google AI) is an other alternative, however evaluating their cost-to-performance ratio remains an open task.

Because I hit a wall on this part of the pipeline, and it became clear that this project is infeasible (at least for the time being), I was discouraged from exploring more options about the other parts of the system. And actually, to be honest, I did not have much time available to dig deeper — this whole attempt was done in a few weeks, and I had hoped to finish this project quickly.

Data quality also presented a major hurdle. The university website heavily relies on scanned PDFs, which complicate text extraction. Furthermore, automatically downloading the site’s entire 10-to-15-year archive runs the risk of feeding outdated policies into the RAG pipeline, diluting answer accuracy. Rather than an automated scraper, a manually curated list of current, official documents is a far safer and more effective approach.

In the short term, utilizing an off-the-shelf solution like NotebookLM (as described on the second section) offers the fastest path to value with minimal overhead. To make this approach effective, I recommend creating role-specific notebooks—such as separate, targeted notebooks for staff and students—rather than a single, monolithic knowledge base.