DIRS is a modular Retrieval-Augmented Generation (RAG) system built in Python. Upload any PDF document and ask natural language questions about its content — entirely on-premises, with no cloud APIs or external data transfer.
DIRS turns unstructured PDF documents into a searchable knowledge base. When a document is uploaded, the system extracts and chunks the text, generates vector embeddings, and stores them in a vector database. When a question is asked, it retrieves the most relevant chunks using a hybrid retrieval strategy — combining semantic vector search with BM25 keyword ranking — and passes them to a local LLM to generate a grounded answer.
This reduces hallucinations and keeps responses tightly tied to the document content.
- PDF ingestion and structured text extraction
- Intelligent text chunking with configurable size and overlap
- Hybrid retrieval combining semantic search (FAISS / Chroma) and keyword ranking (BM25) with weighted score fusion
- In-memory caching of indices and retrievers for reduced query latency
- Choice of embedding models: BGE-small, MiniLM, E5-small
- Choice of local LLMs via Ollama: LLaMA 3, Qwen 2.5, Gemma
- Performance tracking and experiment logging to CSV
- Role-based UI: Admin (index builder) and User (question answering)
- Source transparency — retrieved chunks shown alongside answers
PDF Upload → Text Extraction → Chunking → Embedding → Vector Store
↓
User Question → Query Embedding → Hybrid Retrieval (Vector + BM25)
↓
Context → LLM → Answer
| Component | Technology |
|---|---|
| Language | Python 3.10+ |
| Web Interface | Streamlit |
| LLM Serving | Ollama |
| Language Models | LLaMA 3, Qwen 2.5, Gemma |
| Embeddings | Sentence Transformers (BGE, MiniLM, E5) |
| Vector Search | FAISS, ChromaDB |
| Keyword Search | BM25 (rank-bm25) |
| PDF Parsing | PyPDF |
Admin — Uploads PDFs, selects embedding model and vector database, builds and manages the index.
User — Selects an indexed document, selects an LLM, asks questions, and receives grounded answers with source citations and performance metrics.
For full installation instructions, environment setup, and usage guide, see SETUP.md.
Aman Srivastava amansri345@gmail.com