Skip to content

Ask Chitrank

RAG-powered AI assistant that answers questions about Chitrank using resume, portfolio, LinkedIn, and testimonial data.

Type URL
📚 Documentation Project Documentation
👤 Portfolio About me

What It Does

Ask Chitrank is a conversational AI that answers questions about Chitrank Agnihotri — his experience, projects, skills, and background. It uses RAG to ground every answer in real portfolio data and avoid hallucination.

Example questions it answers:

  • "What projects has Chitrank built?"
  • "What is his tech stack?"
  • "How many years of experience does he have?"
  • "What do his colleagues say about him?"
  • "Has he worked with AI products?"

How It Works

User question
    ↓
Cheap safety pre-router
    ↓ bypass                     ↓ continue
Canned response             Exact cache (Case-insensitive match)
                                ↓ hit                        ↓ miss
                        Return cached response      Embed question (Voyage AI / Local Fallback)
                                                        ↓
                                                Check semantic cache (Similarity > 0.95)
                                                        ↓ hit                        ↓ miss
                                                Return cached response      Expand query & Search chunks
                                                                                ↓
                                                                        Query-aware local reranking
                                                                                ↓
                                                                      Retrieval confidence gate
                                                                                ↓ pass         ↓ fail
                                                                        Build prompt + context  Canned fallback
                                                                                ↓
                                                                        Groq LLM (Llama 4 Scout 17B-16E)
                                                                                ↓
                                                                        Store in cache
                                                                                ↓
                                                                        Stream response

Knowledge Base Shape

Source Content type Retrieval shape
Resume PDF Experience, skills, education Section-aware chunks
Sanity CMS Projects Overview, contribution, and link evidence documents
Testimonials Colleague recommendations One evidence document per testimonial
LinkedIn Profile summary, links, recommendations Compact profile and recommendation evidence documents

Exact chunk counts vary as source content changes and as structured evidence documents are emitted during ingestion.

Recent ROI Improvements

  • LLM Upgrade: Migrated to Llama 4 Scout 17B-16E on Groq for improved reasoning and response quality.
  • Query Expansion: Synonym-based expansion improves retrieval recall for short or ambiguous queries.
  • Provider Fallback: Support for local embeddings (all-MiniLM-L6-v2) provides a resilient alternative to Voyage AI.
  • Structured feature extraction now creates retrieval-friendly evidence documents instead of relying only on broad source blobs.
  • Retrieval now combines vector search with cheap local reranking, improving relevance without increasing provider spend.
  • Chat now uses a cheap pre-router plus a retrieval confidence gate, improving safety and saving tokens on unsupported questions.
  • DEV_MODE now supports fictional seeded data and local fake providers so local API work does not burn Groq or Voyage quota.

Roadmap

  • [x] Phase 1 — Database layer (pgvector, Supabase, Alembic migrations)
  • [x] Phase 2 — Ingestion pipeline (resume PDF, Sanity CMS, LinkedIn)
  • [x] Phase 3 — Retrieval layer (vector search + semantic cache)
  • [x] Phase 4 — Chat layer (prompt engineering + Groq LLM)
  • [x] Phase 5 — FastAPI + streaming (SSE chat endpoint)
  • [x] Phase 7 — Sanity webhook auto-sync
  • [ ] Phase 6 — Frontend chat widget (Next.js)

Known Limitations

  • Groq free tier rate limits still apply in production mode
  • Supabase free tier can pause after inactivity, making the first request slower
  • The semantic cache threshold of 0.95 is intentionally strict and may miss some near-duplicates
  • Local reranking is heuristic-based, so it should still be tuned against real production queries over time
  • Safety pre-routing and retrieval confidence are heuristic-based and should also be tuned against real production traffic

Contributing

This is a personal portfolio project. Issues and PRs welcome.

Developed by Chitrank Agnihotri