OmniRAG is a multi-format knowledge engine that handles PDFs, images, codebases, audio, video, and web pages in one pipeline, delivering questions with exact source citations.

The tech stack blends Python, FastAPI, ChromaDB, PostgreSQL, Redis, Whisper, and optionally Ollama for local AI deployment.

Uploaded files are routed to the appropriate parser, chunked, embedded into a vector store, stored in ChromaDB, while a parallel BM25 index is built to support query-time retrieval; results are merged and re-ranked before sending the top chunks to the LLM.

OmniRAG uses a hybrid search that combines embeddings for semantic understanding with BM25 for precise keyword matching, addressing content that isn’t embedding-friendly like identifiers, code symbols, and exact numbers.

For those who want to try it, there’s a practical invitation with a link to the GitHub repository: github.com/isidhartha/omni-rag.

The system supports multi-turn conversations with memory, ensuring follow-up questions stay in context, and each answer includes citations showing document name and exact page or timestamp.

Inputs include PDFs with full-text extraction and metadata, images with OCR and diagram understanding, code repositories with AST-aware indexing, audio and video transcriptions via Whisper, and web pages scraped into the knowledge base.