Chat With Your Files (RAG Pipeline)

Upload Markdown → ask questions → get grounded answers sourced only from your content.

Snapshot (Why It Matters)

Demonstrates end‑to‑end Retrieval Augmented Generation with:

• Automatic ingestion & processing (triggers + Edge Functions)
• Vector similarity search (pgvector, HNSW)
• Hybrid embeddings (server for documents, client for queries)
• Streaming LLM responses with constrained context
• Strong data isolation via Row Level Security

Architecture Flow

Upload → Storage trigger → process fn (parse + chunk) → sections → embed fn (batch embeddings)
→ Postgres (pgvector index)
→ Query: local embedding (Web Worker) → match_document_sections → OpenAI (stream) → UI

Tech Stack

Next.js (App Router) · Supabase (Auth, Storage, Postgres, Edge Functions) · pgvector (HNSW) · OpenAI API · Transformers.js (in-browser) · TypeScript · Tailwind

Core Data & Retrieval

Object	Purpose
documents	One row per uploaded file
document_sections	Chunked text + embedding (vector 384)
RPC match_document_sections	Inner product similarity lookup

Retrieval: normalized embeddings + thresholded inner product for fast filtering.

Key Decisions

Choice	Benefit
DB triggers over external queue	Fewer moving parts early on
Client-side query embedding	Lower latency & cost control
Heading-based chunking	Keeps semantic cohesion
Streaming responses	Better UX under longer generations

Security

• RLS + storage policies enforce per-user isolation
• Edge Functions require bearer token
• Only selected chunk text goes to OpenAI

Selected Source Files

File	Role
supabase/migrations/*.sql	Schema, triggers, RLS, vector index
supabase/functions/process/index.ts	Parse + chunk markdown
supabase/functions/embed/index.ts	Batch embedding writes
supabase/functions/chat/index.ts	Retrieval + streaming answer
supabase/functions/_lib/markdown-parser.ts	Heading-based chunk logic
lib/hooks/use-pipeline.ts	Web Worker embedding hook
app/chat/page.tsx	Chat UI + streaming integration

Local Setup

npm install
supabase start
npm run dev

Env:

NEXT_PUBLIC_SUPABASE_URL=...
NEXT_PUBLIC_SUPABASE_ANON_KEY=...
OPENAI_API_KEY=...

Regenerate DB types: npm run gen:types

Possible Extensions

Citizens / citations, metadata filtering, observability, rate limiting, tests for chunker & retrieval.

Focus Demonstrated

Applied RAG design, secure multi-tenant data handling, vector retrieval wiring, and pragmatic trade-offs for latency & cost.

---