VerbaMind — grounded document intelligence

VerbaMind is a full-stack Hybrid RAG application for asking verifiable questions over uploaded PDFs. It combines semantic retrieval with PostgreSQL full-text search, fuses results with Reciprocal Rank Fusion (RRF), and returns the source context alongside every answer.

Why this is a portfolio project

• Hybrid retrieval: Gemini embeddings provide semantic retrieval; PostgreSQL FTS catches exact terms, names, and identifiers.
• RRF fusion: Dense and keyword rankings are blended rather than relying on one retrieval signal.
• Parent-child retrieval: small, overlapping child chunks are embedded for precision; their parent paragraph is supplied for useful answer context.
• Grounding: the generator is instructed to answer only from retrieved sources, cite [Source N], and abstain when evidence is missing.
• Inspectable outputs: the UI shows every ranked source block that was supplied to the LLM.
• Production-aware design: secrets and database URLs use environment variables; upload size/type validation is enforced server-side.

Architecture

PDF upload → PDF.js extraction → overlapping child chunks → Gemini embeddings
                                              ↓
                                      PostgreSQL + pgvector
                                              ↓
Question → vector search + PostgreSQL FTS → RRF → source-labelled context → Groq

Stack

Next.js App Router · TypeScript · PostgreSQL + pgvector · Drizzle ORM · Gemini embeddings · Groq generation · LangGraph · LangSmith (optional) · Tailwind CSS

Run locally

1. Copy .env.example to .env.local and set GEMINI_API_KEY and GROQ_API_KEY. Do not commit this file.

2. Start local Postgres with pgvector:

   docker compose up -d

3. Install packages and create the schema:

   pnpm install
   pnpm db:push
   pnpm dev

4. Open http://localhost:3000, upload a PDF (up to 10 MB), and ask a question.

Deploy

Use Vercel for the Next.js application and a hosted PostgreSQL database with the vector extension enabled (for example Neon, Supabase, or Railway).

Set these environment variables in the deployment project:

DATABASE_URL=postgresql://...
GEMINI_API_KEY=...
GROQ_API_KEY=...
LANGCHAIN_TRACING_V2=true        # optional
LANGCHAIN_API_KEY=...            # optional
LANGCHAIN_PROJECT=verbamind      # optional

Run pnpm db:push against the production DATABASE_URL once, then deploy the application. The build command is pnpm build and the start command is pnpm start.

Production notes

• MemorySaver preserves conversation state only for the active server instance. Move chat history to Redis or Postgres before claiming durable multi-user memory.
• Add authentication and a persistent rate limiter before making uploads public; LLM API endpoints otherwise incur costs for anyone who reaches them.
• Store documents per user/tenant and filter retrieval by owner before using private data.
• Build a small labelled question set and measure retrieval recall and grounded-answer quality before tuning thresholds.

Verification

pnpm typecheck
pnpm build