Mini-RAG
Mini-RAG: Citation-Grounded Retrieval-Augmented Generation
Ask your documents. Get answers with inline citations.

Abstract preview - not a screenshot
01 / Problem
Upload PDF or text, ask questions, and get answers grounded in retrieved evidence with inline citations.
02 / Approach
- 01Character-based chunking with metadata, embedded with Gemini text-embedding-004.
- 02Qdrant Cloud cosine top-k search, then lightweight reranking that filters low-relevance chunks.
- 03Grounded generation with inline citations to the retrieved evidence.
- 04FastAPI backend on Render, frontend on Vercel, credentials and vector-DB ops kept server-side.
03 / Architecture
04 / Key details
- Pipeline: ingestion → text extraction → character-based chunking → metadata → Gemini text-embedding-004 embeddings → Qdrant Cloud cosine top-k → lightweight reranking → grounded generation.
- Reranking filters low-relevance chunks before generation.
- FastAPI backend on Render, frontend on Vercel, credentials and vector-DB ops kept server-side.
05 / Results & signals
- top-k
- Qdrant cosine search
- [1][2]
- Inline citations
- Rerank
- Low-relevance filter
- Live
- Render + Vercel
06 / Links
Next entry →
AXON
A voice- and text-controlled desktop assistant with dual-LLM routing.