PDF → Vector Database · LangChain Edition
Load a PDF. Embed it. Search it semantically. Ask questions in any language.
VecRAG is a local Retrieval-Augmented Generation workbench. Drop in any PDF, choose an embedding strategy, embed the text into a local ChromaDB vector store, then search or ask questions — all running on your own machine with no cloud API keys required.
Built for Mongolian and English documents, with multilingual embedding models selected by default.
| Feature | Details |
|---|---|
| 📄 PDF Loading | PyMuPDF + PyPDF fallback, page-by-page preview with page renderer |
| ✂ 6 Chunking Strategies | Recursive, Semantic, Sentence, Paragraph, Token, Markdown |
| 🔍 Chunk Preview | Visual quality bar, per-chunk badges (GOOD / TOO SHORT / TOO LONG) before embedding |
| 🧠 Multilingual Embeddings | HuggingFace local models or Ollama — Mongolian + English optimised |
| 💾 ChromaDB | Persistent local vector store via LangChain Chroma |
| 🔎 3 Search Modes | Semantic similarity, MMR (diverse), Score-threshold |
| 🦙 LLM QA | Ask questions answered by any Ollama model (LFM 2.5 Thinking default) |
| ⚡ Non-blocking UI | Embed, Search, and Ask all run in background threads — UI never freezes |
| 📊 Live Logs | Terminal-style log for every background operation |
| 🌐 Bilingual | Mongolian + English prompts, auto-detect answer language |
|
Documents tab — upload, page preview, chunk preview with quality analysis |
Embed tab — live terminal log while embedding runs in background |
|
Search tab — semantic / MMR / threshold search with live log |
Ask tab — LLM answers with source chunk citations |
📁 Put your own screenshots in
images/and update the table above.
vecrag/
├── app.py # Entrypoint — sidebar config + tab router
├── requirements.txt
├── images/
│ └── demo.gif
│
├── .streamlit/
│ └── config.toml # 500MB upload limit
│
├── core/
│ ├── loader.py # LangChain PDF loading (PyMuPDF + PyPDF)
│ ├── splitter.py # 6 LangChain text splitter strategies
│ ├── embedder.py # HuggingFace / Ollama embedding factories
│ ├── vectorstore.py # ChromaDB wrapper + background embed/search workers
│ └── chain.py # LCEL RAG chain + background LLM worker
│
├── panels/
│ ├── documents.py # Tab 1 — Upload, page viewer, chunk preview
│ ├── embed.py # Tab 2 — Embed config + live log
│ ├── search.py # Tab 3 — Vector search + live log
│ └── ask.py # Tab 4 — LLM QA + source chunks
│
└── ui/
├── styles.py # Navy + teal dark theme CSS
└── components.py # Reusable HTML helpers (cards, terminal, stat rows)
- Python 3.10+
- Ollama installed and running
git clone https://github.com/tsejavhaa/vecrag.git
cd vecrag
pip install -r requirements.txt# Default LLM for the Ask tab
ollama pull lfm2.5-thinking
# Optional — better Mongolian quality
ollama pull qwen2.5:3bstreamlit run app.pyOpen http://localhost:8501 in your browser.
Upload PDF → Preview chunks → Embed → Search → Ask
Tab 1 Tab 1 Tab 2 Tab 3 Tab 4
- Documents — Upload your PDF. Browse pages. Click Run Chunk Preview to see how your chosen strategy splits the text before committing to a full embed.
- Embed — Press Embed and watch the live terminal. Runs in background so the UI stays responsive.
- Search — Type a query. Choose a search mode. Results show with relevance scores and full chunk text.
- Ask — Type a question in Mongolian or English. The LLM reads the top-k retrieved chunks and writes a synthesised answer with citations.
| Strategy | Tag | Best for |
|---|---|---|
| Recursive Character | GENERAL | Most PDFs — good default |
| Semantic ✦ | BEST QUALITY | Best retrieval quality — splits at topic boundaries |
| Sentence Transformer Tokens | TOKEN-AWARE | Never exceeds model context window |
| Paragraph | STRUCTURE | Well-formatted PDFs with clear paragraphs |
| Token (tiktoken) | LLM-READY | When feeding chunks directly to a GPT-style LLM |
| Markdown Header | STRUCTURED | Headed / structured documents |
Tip: Use the Chunk Preview in the Documents tab to compare strategies on your actual document before embedding.
All models below run locally — no API key needed.
| Model | Dims | Languages | Notes |
|---|---|---|---|
paraphrase-multilingual-MiniLM-L12-v2 |
384 | 50+ incl. Mongolian | ⭐ Default — great MN+EN balance |
intfloat/multilingual-e5-small |
384 | 100 | Strong retrieval benchmarks |
intfloat/multilingual-e5-base |
768 | 100 | Best quality — needs more RAM |
all-MiniLM-L6-v2 |
384 | English only | Fast fallback for EN-only docs |
You can also use Ollama embeddings (nomic-embed-text, mxbai-embed-large) by switching the provider in the sidebar.
All models run via Ollama — fully local, no internet required after pulling.
| Model | Size | Notes |
|---|---|---|
lfm2.5-thinking |
~2 GB | ⭐ Default — fast reasoning model |
qwen2.5:3b |
~2 GB | Excellent Mongolian / Cyrillic quality |
llama3.2:3b |
~2 GB | Good general Mongolian + English |
llama3.2:1b |
~1 GB | Lightest — CPU-friendly |
qwen2.5:7b |
~5 GB | Best multilingual quality |
Any model in your ollama list works — type it directly in the sidebar model field.
The sidebar shows live status for every pipeline stage:
● PDF angels_and_demons_mn.pdf
● Vector DB 76 chunks
● Embed done
● Search 8 results
● Ask done 4s
Dot colours: 🟢 green = OK · 🟡 yellow = warning / changed · 🔵 pulsing = running · ⚫ off = idle
All settings live in the sidebar — no config files to edit.
| Section | Settings |
|---|---|
| Text Splitter | Strategy, Chunk size (chars), Overlap (chars) |
| Embedding Model | Provider (HuggingFace / Ollama), Model, Device (cpu / cuda / mps) |
| LLM | Model (any Ollama model), Ollama URL |
| Storage | Collection name, Persist directory |
streamlit>=1.30.0
langchain>=0.3.0
langchain-core>=0.3.0
langchain-text-splitters>=0.2.0
langchain-community>=0.2.0
langchain-huggingface>=0.0.3
langchain-ollama>=0.1.0
langchain-chroma>=0.1.0
langchain-experimental>=0.0.60
chromadb>=0.5.0
sentence-transformers>=3.0.0
pymupdf>=1.23.0
pypdf>=4.0.0
requests>=2.31.0
MIT — free to use, modify, and distribute.