Mini-RAG

Mini-RAG: Citation-Grounded Retrieval-Augmented Generation

Ask your documents. Get answers with inline citations.

Category
AI Systems · Retrieval
  • Gemini
  • Qdrant Cloud
  • FastAPI
  • Render
  • Vercel

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

  1. 01Character-based chunking with metadata, embedded with Gemini text-embedding-004.
  2. 02Qdrant Cloud cosine top-k search, then lightweight reranking that filters low-relevance chunks.
  3. 03Grounded generation with inline citations to the retrieved evidence.
  4. 04FastAPI backend on Render, frontend on Vercel, credentials and vector-DB ops kept server-side.

03 / Architecture

Mini-RAG system architectureUpload, then Extract & chunk, then Embed, then Qdrant Cloud, then Retrieve top-k, then Rerank, then Grounded answer.SERVER-SIDE · RENDER01UploadPDF OR PLAIN TEXT02Extract & chunkCHARACTER-BASED CHUNKS + METADATA03EmbedGEMINI TEXT-EMBEDDING-00404Qdrant CloudCOSINE VECTOR INDEX05Retrieve top-kQUERY EMBEDDED, COSINE SEARCHQUERY06RerankDROP LOW-RELEVANCE CHUNKS07Grounded answerGEMINI · INLINE CITATIONS [1][2]
Fig. 01 - System architectureSimplified

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

Next entry →

AXON

A voice- and text-controlled desktop assistant with dual-LLM routing.