Computer Vision (CV) Papers RAG - Multimodal Research Assistant

A Retrieval-Augmented Generation (RAG) system for searching and querying Computer Vision research papers. Ask questions in natural language and get AI-synthesized answers with sources and architecture figures.

Features

Automatic Paper Download: Fetch papers from arXiv by topic
Semantic Search: Find relevant papers using natural language queries (not just keywords)
AI-Powered Answers: Get synthesized responses from retrieved passages using GPT-4o
Architecture Figures: Search and display architecture diagrams from papers
Multimodal Search: Combine text and figure search for comprehensive results
Source Citations: Every answer includes paper sources with similarity scores
Modern Web Interface: Clean UI with Ask AI and Search modes
REST API: Full API with Swagger documentation
Smart Filtering: Automatically filters out references and related work sections

Demo

Ask Mode (Web Interface)

Ask a question and get a synthesized answer:

Q: "How does person re-identification work with gait features?"

A: "Person re-identification using gait features combines appearance 
   and motion patterns. GAF-Net extracts gait features from skeletal 
   data using Graph Convolutional Networks (GCN) and fuses them with 
   RGB appearance features. This multimodal approach achieves 89.6% 
   Rank-1 accuracy on iLIDS-VID dataset..."

Sources: 
  - VISAPP_2024_157_CR.pdf (68.4% match)
  - algorithms-17-00352-v2.pdf (58.5% match)

Related Figures:
  - [GAF-Net Architecture Diagram]

Command Line Mode

# Interactive RAG
$ python src/rag.py
RAG System for Computer Vision Papers
Your question: How does attention work in transformers?
ANSWER: [Generated answer with sources]

# Multimodal search
$ python src/multimodal_rag.py
Search: transformer encoder decoder
[Results with text chunks + architecture diagrams]

How It Works

RAG Pipeline

+------------------------------------------------------------------+
|                         USER QUESTION                            |
|           "How does person re-identification work?"              |
+------------------------------------------------------------------+
                              |
                              v
+------------------------------------------------------------------+
|                    1. RETRIEVAL (Local)                          |
|  +-------------+    +-------------+    +-------------+           |
|  |  Embedding  | -> |  ChromaDB   | -> |   Top-K     |           |
|  |  (MiniLM)   |    |   Search    |    |  Passages   |           |
|  +-------------+    +-------------+    +-------------+           |
|        FREE              FREE             FREE                   |
+------------------------------------------------------------------+
                              |
                              v
+------------------------------------------------------------------+
|                    2. GENERATION (LLM)                           |
|  +-------------+    +-------------+    +-------------+           |
|  |  Passages   | -> |   GPT-4o    | -> | Synthesized |           |
|  |  + Question |    |   mini      |    |   Answer    |           |
|  +-------------+    +-------------+    +-------------+           |
|                        ~$0.001/request                           |
+------------------------------------------------------------------+
                              |
                              v
+------------------------------------------------------------------+
|              RESPONSE: Answer + Sources + Figures                |
+------------------------------------------------------------------+

Key Concepts

Component	What it does	Technology
Embeddings	Convert text to vectors for similarity search	all-MiniLM-L6-v2 (local, free)
Vector DB	Store and search embeddings efficiently	ChromaDB
Retrieval	Find relevant passages from papers	Cosine similarity
Generation	Synthesize answer from passages	GPT-4o-mini
Figures	Extract and describe architecture diagrams	PyMuPDF + GPT-4o Vision

Project Structure

rag-cv-papers/
│
├── src/                           # Core application modules
│   ├── api.py                     # FastAPI REST API (/ask, /search endpoints)
│   ├── arxiv_fetcher.py          # Download papers from arXiv automatically
│   ├── embeddings.py             # Text chunking and indexing into ChromaDB
│   ├── document_processor.py     # PDF processing, intelligent section filtering
│   ├── figure_extractor.py       # Extract architecture figures from PDFs
│   ├── index_figures.py          # Index figure descriptions into ChromaDB
│   ├── rag.py                    # Core RAG pipeline (retrieval + generation)
│   ├── multimodal_rag.py         # Multimodal search (text + figures)
│   └── test_llm.py               # Test OpenAI API connection
│
├── check_desc.py                  # Validate figure descriptions quality
├── debug_figures.py               # Debug ChromaDB figure collection
├── debug_p5.py                    # Debug specific figure extraction issues
├── fix_descriptions.py            # Re-generate invalid figure descriptions
│
├── static/
│   └── index.html                 # Web interface
│
├── data/
│   ├── papers/                    # Your PDF papers (input)
│   ├── figures/                   # Extracted figures (generated)
│   └── figure_descriptions.json   # Figure metadata (generated)
│
├── chroma_db/                     # Vector database (generated)
│   ├── cv_papers/                 # Text chunks collection
│   └── cv_figures/                # Figure descriptions collection
│
├── .env                           # API keys (OPENAI_API_KEY)
├── .env.example                   # Template for environment variables
├── .gitignore                     # Git ignore rules
├── requirements.txt               # Python dependencies
└── README.md                      # This file

Installation

Prerequisites

Python 3.10 or higher
OpenAI API key (for Ask mode and figure descriptions)

Step 1: Clone the repository

git clone https://github.com/Moncef-Bj/rag-cv-papers.git
cd rag-cv-papers

Step 2: Create a virtual environment

# Windows
python -m venv .venv
.venv\Scripts\activate

# Linux/Mac
python -m venv .venv
source .venv/bin/activate

Step 3: Install dependencies

pip install -r requirements.txt

Step 4: Configure API key

Create a .env file (copy from .env.example):

cp .env.example .env

Edit .env and add your OpenAI API key:

OPENAI_API_KEY=sk-your-api-key-here

Step 5: Test the setup

# Test OpenAI connection
python src/test_llm.py

If you see "API key loaded successfully" and a response from GPT, you're ready!

Complete Workflow

Here's a complete example starting from scratch with "person re-identification" topic:

1. Download Papers from arXiv

# Download 10 papers on person re-identification
python src/arxiv_fetcher.py "person re-identification"

# Or download papers on multiple topics (default behavior)
python src/arxiv_fetcher.py

Output:

Searching arXiv for: person re-identification
Download folder: data/papers
--------------------------------------------------
Deep Learning for Person Re-identification: A Survey...
   Authors: Liang Zheng, Yi Yang, Alexander G. Hauptmann
   Published: 2016-10-13
   Downloaded: 1610.02984_Deep_Learning_for_Person_Re-identification.pdf

... (9 more papers)

==================================================
Fetched 10 papers
Next: Run 'python src/embeddings.py' to index new papers

2. Index Text Chunks

python src/embeddings.py

What happens:

Loads all PDFs from data/papers/
Extracts text and filters out references, related work
Chunks documents into ~1000 character pieces
Generates embeddings (vectors) for each chunk
Stores everything in ChromaDB

Output:

============================================================
STEP 1: Loading and chunking PDFs
============================================================
Found 10 PDF files
--------------------------------------------------
Processing: 1610.02984_Deep_Learning_for_Person_Re...
   Pages: 23
   Chunks: 47 -> 31 (filtered 16)
...

PROCESSING SUMMARY
============================================================
Total chunks processed:    458
Chunks kept:               312
Filtered (references):     89
Filtered (related work):   38
Filtered (low quality):    19
Filter rate:               31.9%
============================================================

STEP 2: Creating vector store
============================================================
Vector store created!
   Total vectors: 312

3. Extract and Describe Figures (Optional but Recommended)

# Extract architecture diagrams from PDFs
python src/figure_extractor.py

What happens:

Scans PDFs for figures (pages 1-10)
Filters by size, aspect ratio
Saves to data/figures/
Generates descriptions using GPT-4o Vision
Saves metadata to data/figure_descriptions.json

Cost: ~$0.01 per figure

Output:

============================================================
EXTRACTING ARCHITECTURE FIGURES FROM CV PAPERS
============================================================
Processing: 1610.02984_Deep_Learning_for_Person_Re...
   Found 3 candidate figures
   Saved: 1610.02984_Deep_Learning_p3_img1.png
   Description: "A CNN architecture showing ResNet-50 backbone..."
...

SUMMARY
============================================================
Papers processed: 10
Figures extracted: 27
Descriptions generated: 27

4. Index Figures for Search

python src/index_figures.py

Output:

============================================================
 INDEXING FIGURES INTO CHROMADB
============================================================
   Total figures: 27
   Valid descriptions: 25
Indexed 25 figures!
   Collection: cv_figures

5. Verify Everything Works

# Check figure descriptions quality
python check_desc.py

# Verify what's in ChromaDB
python debug_figures.py

6. Start Using!

Option A: Web Interface

python -m uvicorn src.api:app --reload --port 8000
# Open http://localhost:8000

Option B: Command Line RAG

python src/rag.py

Option C: Multimodal Search

python src/multimodal_rag.py

Usage

Web Interface

Start the server:

python -m uvicorn src.api:app --reload --port 8000

Open http://localhost:8000 in your browser
Choose mode:
- Ask AI: Get synthesized answers with sources
- Search: Get raw search results
Example questions:
- "How does person re-identification work?"
- "What are the main challenges in multi-camera tracking?"
- "Explain the difference between CNN and Transformer architectures"
- "What loss functions are used for metric learning?"

Command Line Tools

1. Interactive RAG (Ask Questions)

python src/rag.py

Example session:

============================================================
RAG System for Computer Vision Papers
============================================================
Ask questions about the papers in your database.
Type 'quit' to exit.

 Your question: How does person re-identification work?

Retrieving context for: How does person re-identification work?
Context retrieved (3421 chars)
Generating answer...

ANSWER:
Person re-identification (Re-ID) aims to match pedestrian images across 
non-overlapping camera views. Modern approaches use deep learning with 
metric learning objectives. The typical pipeline involves:

1. Feature Extraction: CNN or Transformer backbones extract appearance features
2. Metric Learning: Triplet loss or contrastive loss learn discriminative features
3. Matching: Compute similarity between query and gallery features

According to the papers, state-of-the-art methods achieve 89-95% Rank-1 
accuracy on standard benchmarks like Market-1501 and DukeMTMC-reID.
[Source: 1610.02984_Deep_Learning_for_Person_Re-identification.pdf, Page 4-6]

 Your question: quit
Goodbye!

2. Multimodal Search (Text + Figures)

python src/multimodal_rag.py

Example session:

======================================================================
ARCHITECTURE SEARCH ENGINE
======================================================================
Search for CV architectures using natural language.

Example queries:
  - transformer encoder decoder
  - object detection CNN backbone
  - multi-camera tracking
  - attention mechanism for video

Search: person re-identification CNN architecture

======================================================================
SEARCH: "person re-identification CNN architecture"
======================================================================
   Found: 5 text chunks, 3 figures

----------------------------------------------------------------------
COMBINED RESULTS (sorted by relevance)
----------------------------------------------------------------------

[1] Text | Similarity: 0.842
    Source: 1610.02984_Deep_Learning_for_Person_Re-identificatio...
    Page: 5
    Content: Deep convolutional networks for person Re-ID typically use 
             ResNet or DenseNet backbones. The network learns a metric...

[2] Figure | Similarity: 0.791
    Source: algorithms-17-00352-v2.pdf
    Page: 8
    Title: GAF-Net Architecture Overview
    Description: The architecture shows a dual-branch network with CNN 
                 for appearance and GCN for gait features...
    File: algorithms-17-00352-v2_p8_img2.png

[3] Text | Similarity: 0.756
    Source: VISAPP_2024_157_CR.pdf
    Page: 3
    Content: The proposed architecture uses a ResNet-50 backbone 
             pretrained on ImageNet, followed by a global average...

----------------------------------------------------------------------
ARCHITECTURE FIGURES
----------------------------------------------------------------------

   [1] algorithms-17-00352-v2_p8_img2.png
       Paper: algorithms-17-00352-v2.pdf
       Path: data/figures/algorithms-17-00352-v2_p8_img2.png
       Title: GAF-Net Architecture Overview

   [2] VISAPP_2024_157_CR_p5_img4.png
       Paper: VISAPP_2024_157_CR.pdf
       Path: data/figures/VISAPP_2024_157_CR_p5_img4.png
       Title: Feature Fusion Module

   [3] 1610.02984_Deep_Learning_p7_img1.png
       Paper: 1610.02984_Deep_Learning_for_Person_Re-identification.pdf
       Path: data/figures/1610.02984_Deep_Learning_p7_img1.png

3. Download Papers from arXiv

Single topic:

python src/arxiv_fetcher.py "person re-identification"

Multiple topics (default):

python src/arxiv_fetcher.py

Suggested Topics for CV Research:

# Person Re-identification & Tracking
"person re-identification"
"multi-camera tracking"
"pedestrian detection"
"gait recognition"

# Human Pose & Activity
"human pose estimation"
"action recognition"
"3D human reconstruction"

# Object Detection & Segmentation
"object detection transformer"
"instance segmentation"
"semantic segmentation"

# Video Understanding
"video object tracking"
"temporal action detection"
"video captioning"

# 3D Vision
"visual SLAM"
"3D object detection"
"depth estimation"

# Transformers & Attention
"vision transformer"
"self-attention mechanism"
"cross-attention fusion"

# Few-shot & Meta-learning
"few-shot learning"
"meta-learning computer vision"

# Domain Adaptation
"domain adaptation"
"transfer learning"

Custom download example:

# Download 20 papers on transformers for CV
python src/arxiv_fetcher.py "vision transformer" --max-results 20

API Endpoints

Endpoint	Method	Description
`/`	GET	Web interface
`/health`	GET	Health check
`/stats`	GET	Collection statistics
`/search`	GET	Semantic search (retrieval only)
`/ask`	GET	Ask AI (RAG with LLM synthesis)
`/figures/{name}`	GET	Serve figure images
`/docs`	GET	Swagger API documentation

Example API Calls

Health check:

curl http://localhost:8000/health

Get stats:

curl http://localhost:8000/stats

Response:

{
  "text_chunks": 312,
  "figures": 25,
  "papers": 10
}

Search (retrieval only):

curl "http://localhost:8000/search?q=person+re-identification&n_text=5&n_figures=3"

Ask AI (full RAG):

curl "http://localhost:8000/ask?q=How+does+person+re-identification+work"

Response:

{
  "question": "How does person re-identification work?",
  "answer": "Person re-identification (Re-ID) is the task of matching...",
  "sources": [
    {
      "paper": "1610.02984_Deep_Learning_for_Person_Re-identification.pdf",
      "content": "Person re-identification aims to match pedestrians...",
      "similarity": 0.842,
      "page": 5
    }
  ],
  "figures": [
    {
      "filename": "algorithms-17-00352-v2_p8_img2.png",
      "title": "GAF-Net Architecture",
      "description": "Dual-branch network combining appearance and gait..."
    }
  ]
}

Downloading Papers from arXiv

The arxiv_fetcher.py tool automatically downloads research papers from arXiv.

Basic Usage

# Single topic
python src/arxiv_fetcher.py "person re-identification"

# Multiple topics (default config)
python src/arxiv_fetcher.py

Default Topics

When run without arguments, it downloads papers on these topics:

1. person re-identification (15 papers)
2. multi-camera tracking (15 papers)
3. pedestrian detection (15 papers)
4. gait recognition (15 papers)
5. human pose estimation (15 papers)
6. object detection transformer (15 papers)
7. video object tracking (15 papers)
8. action recognition (15 papers)
9. 3D human reconstruction (15 papers)
10. visual SLAM (15 papers)

Total: 150 papers

Customization

Edit src/arxiv_fetcher.py to customize:

# Change topics
topics = [
    "your topic 1",
    "your topic 2",
]

# Change papers per topic
papers = fetch_multiple_topics(topics=topics, papers_per_topic=20)

Output

Papers are saved to data/papers/ with filenames like:

2312.12345_Paper_Title_Here.pdf

Maintenance & Debugging

Debug Scripts

1. Check Figure Descriptions Quality

python check_desc.py

Validates that all figure descriptions are valid (not GPT refusals).

Output:

============================================================
CHECKING VISAPP AND ALGORITHM FIGURES
============================================================

Filename: VISAPP_2024_157_CR_p5_img4.png
Source: VISAPP_2024_157_CR.pdf
Description length: 234 chars
Description: The architecture diagram shows a feature fusion module 
             that combines appearance and gait features using...
STATUS: Description looks OK!

2. Debug ChromaDB Contents

python debug_figures.py

Shows what's actually indexed in ChromaDB.

Output:

============================================================
FIGURES IN CHROMADB
============================================================
  algorithms-17-00352-v2.pdf           -> algorithms-17-00352-v2_p8_img2.png
  VISAPP_2024_157_CR.pdf               -> VISAPP_2024_157_CR_p5_img4.png
  1610.02984_Deep_Learning_for_Per...  -> 1610.02984_Deep_Learning_p7_img1.png

============================================================
SEARCHING FOR VISAPP AND ALGORITHM FIGURES
============================================================

FOUND:
  source_pdf: VISAPP_2024_157_CR.pdf
  filename: VISAPP_2024_157_CR_p5_img4.png
  title: Feature Fusion Architecture

3. Debug Specific Figure Issues

python debug_p5.py

Detailed debugging for why a specific figure wasn't extracted/indexed.

4. Fix Invalid Descriptions

python fix_descriptions.py

Re-generates descriptions for figures that have invalid descriptions (GPT refusals).

Output:

============================================================
RE-DESCRIBING INVALID FIGURES
============================================================

Re-describing: paper_p3_img2.png
  Old description: I cannot analyze this image...
  New description: The diagram shows a ResNet architecture with skip...
  STATUS: Success!

============================================================
Updated 3 figures
Saved to data/figure_descriptions.json

Now run: python src/index_figures.py

5. Test LLM Connection

python src/test_llm.py

Verifies that your OpenAI API key works correctly.

Configuration

Document Processing

The system intelligently filters paper sections:

Kept	Filtered
Abstract	References
Introduction	Related Work
Method/Approach	Acknowledgments
Architecture	Appendix
Experiments	Bibliography
Results
Conclusion

This filtering reduces noise and improves search quality by ~30%.

Figure Extraction Settings

In src/figure_extractor.py:

MIN_IMAGE_SIZE = 20000      # Minimum 20KB (filters out small icons)
MIN_WIDTH = 300             # Minimum width in pixels
MIN_HEIGHT = 200            # Minimum height in pixels
MIN_ASPECT_RATIO = 0.8      # Accept vertical/square figures
TARGET_PAGES = (1, 10)      # Extract from pages 1-10 only
MAX_FIGURES_PER_PAPER = 3   # Max figures per paper
MAX_TOTAL_FIGURES = 100     # Total limit across all papers

Embedding Settings

In src/embeddings.py and src/index_figures.py:

USE_OPENAI = False  # Set to True to use OpenAI embeddings (paid)
                    # False = use local all-MiniLM-L6-v2 (free)

Trade-offs:

Model	Cost	Quality	Speed
all-MiniLM-L6-v2 (local)	FREE	Good	Fast
OpenAI text-embedding-ada-002	~$0.10/1M tokens	Better	Slower

For most use cases, the free local model is sufficient.

Tech Stack

Component	Technology	Purpose
Backend	FastAPI	REST API framework
Vector DB	ChromaDB	Store and search embeddings
Embeddings	all-MiniLM-L6-v2	Convert text to vectors (local, free)
LLM	OpenAI GPT-4o-mini	Answer synthesis
Vision LLM	OpenAI GPT-4o	Figure descriptions
PDF Processing	PyMuPDF (fitz)	Extract text and images
Text Splitting	LangChain	Chunk documents
arXiv API	arxiv (Python)	Download papers
Frontend	HTML/CSS/JavaScript	Web interface

Why These Choices?

ChromaDB: Simple, embedded vector DB - no server needed
MiniLM: Free, local embeddings - no API costs for search
GPT-4o-mini: Good quality at low cost (~$0.15/1M tokens)
FastAPI: Modern, fast, automatic API docs
PyMuPDF: Fast PDF processing with image extraction

Performance & Costs

Performance Metrics

Metric	Value
Indexing speed	~100 papers/minute
Search latency	< 100ms
Ask latency	2-5 seconds (LLM dependent)
Storage	~10MB per 100 papers
RAM usage	~500MB with 100 papers indexed

Costs Breakdown

Operation	Cost	Frequency
Text indexing	FREE	One-time per paper
Search	FREE	Every search
Ask AI	~$0.001	Per question
Figure extraction	~$0.01	One-time per figure
Figure description	~$0.01	One-time per figure

Example Cost for 100 papers:

Text indexing: $0 (free, local)
30 figures @ $0.01 each: $0.30
100 questions @ $0.001 each: $0.10
Total: < $1 for complete setup + 100 questions

Optimization Tips

Use local embeddings (already default) - saves 100% on search costs
Limit figure extraction to key papers - reduce figure description costs
Use GPT-4o-mini (not GPT-4o) for answers - 10x cheaper
Cache results - avoid re-asking same questions

Troubleshooting

"No module named 'chromadb'"

pip install -r requirements.txt

"OPENAI_API_KEY not found"

Create .env file with your API key:

echo "OPENAI_API_KEY=sk-your-key-here" > .env

"No papers found in data/papers/"

Download papers first:

python src/arxiv_fetcher.py "your topic"

Or manually add PDFs to data/papers/.

"Collection 'cv_papers' not found"

Index your papers first:

python src/embeddings.py

Figures not showing in search results

Run figure extraction and indexing:

python src/figure_extractor.py
python src/index_figures.py

"Invalid description" errors

Re-generate descriptions:

python fix_descriptions.py
python src/index_figures.py

ChromaDB is empty after restart

ChromaDB is persistent. If it's empty, you need to re-run:

python src/embeddings.py
python src/index_figures.py

High OpenAI API costs

Set USE_OPENAI = False in embeddings.py (use free local embeddings)
Use GPT-4o-mini instead of GPT-4o
Reduce number of retrieved chunks: TOP_K = 3 in rag.py

Search returns irrelevant results

Check if paper sections are properly filtered:
```
python src/document_processor.py
```

Increase chunk overlap for better context:

# In document_processor.py
chunk_overlap = 300  # Increase from 200

Use more specific queries

Future Improvements

Coming Soon

LaTeX equation rendering - Proper mathematical notation in answers
Docker containerization - One-command deployment

Planned

PDF upload via web interface - No command line needed
Paper summarization - Auto-generate TL;DR for each paper
CLIP embeddings - Visual similarity search for figures

Under Consideration

Citation graph visualization - Explore paper relationships
Automatic paper recommendations - Based on reading history
Streamlit dashboard - Advanced analytics

Ideas

Export to Notion/Obsidian - Note-taking integrations
Collaborative annotations - Highlight and share insights

Contributing

Contributions are welcome! Please:

Fork the repository
Create a feature branch (git checkout -b feature/amazing-feature)
Commit your changes (git commit -m 'Add amazing feature')
Push to the branch (git push origin feature/amazing-feature)
Open a Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

LangChain for document processing utilities
ChromaDB for simple and powerful vector storage
FastAPI for the excellent API framework
OpenAI for GPT models and embeddings
Sentence Transformers for free local embeddings
arXiv for open access to research papers

Citation

If you use this project in your research, please cite:

@software{boujou2025cvrag,
  author = {Boujou, Moncef},
  title = {CV Papers RAG: Multimodal Research Assistant for Computer Vision},
  year = {2025},
  url = {https://github.com/yourusername/rag-cv-papers}
}

Star History

If you find this project useful, please consider giving it a star!

Built with care by a CV researcher, for CV researchers

Name		Name	Last commit message	Last commit date
data/papers		data/papers
src		src
static		static
.env.example		.env.example
.gitignore		.gitignore
README.md		README.md
check_desc.py		check_desc.py
debug_figures.py		debug_figures.py
debug_p5.py		debug_p5.py
fix_descriptions.py		fix_descriptions.py
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

Computer Vision (CV) Papers RAG - Multimodal Research Assistant

Table of Contents

Features

Demo

Ask Mode (Web Interface)

Command Line Mode

How It Works

RAG Pipeline

Key Concepts

Project Structure

Installation

Prerequisites

Step 1: Clone the repository

Step 2: Create a virtual environment

Step 3: Install dependencies

Step 4: Configure API key

Step 5: Test the setup

Complete Workflow

1. Download Papers from arXiv

2. Index Text Chunks

3. Extract and Describe Figures (Optional but Recommended)

4. Index Figures for Search

5. Verify Everything Works

6. Start Using!

Usage

Web Interface

Command Line Tools

1. Interactive RAG (Ask Questions)

2. Multimodal Search (Text + Figures)

3. Download Papers from arXiv

API Endpoints

Example API Calls

Downloading Papers from arXiv

Basic Usage

Default Topics

Customization

Output

Maintenance & Debugging

Debug Scripts

1. Check Figure Descriptions Quality

2. Debug ChromaDB Contents

3. Debug Specific Figure Issues

4. Fix Invalid Descriptions

5. Test LLM Connection

Configuration

Document Processing

Figure Extraction Settings

Embedding Settings

Tech Stack

Why These Choices?

Performance & Costs

Performance Metrics

Costs Breakdown

Optimization Tips

Troubleshooting

"No module named 'chromadb'"

"OPENAI_API_KEY not found"

"No papers found in data/papers/"

"Collection 'cv_papers' not found"

Figures not showing in search results

"Invalid description" errors

ChromaDB is empty after restart

High OpenAI API costs

Search returns irrelevant results

Future Improvements

Coming Soon

Planned

Under Consideration

Ideas

Contributing

License

Acknowledgments

Citation

Star History

About

Topics