What it is

Cosimi is the engine behind this chat - the one you're using right now, live in production. A deterministic GraphRAG retrieval SDK with no LLM in the request path. You ask a question, it embeds it, runs a vector search over a graph of chunks and Q&A pairs, and returns a ranked answer. Same question, same data, same answer - every time. No hallucination, no token cost per reply.

This whole site is the proof. Every answer you get here is Cosimi, not a language model improvising.

How an answer is built

Two layers, both deterministic:

• Curated first. Common questions and the suggestion chips hit a hand-written answer map - exact, on-message, instant.
• Retrieval for the rest. Anything else embeds the query with bge-m3, seeds from the nearest chunks and pairs, walks the graph a couple of hops for context, then ranks. Pure vector math and a graph walk - the model never sees your question.

How it's fed

A separate offline step turns source docs into the corpus: an LLM chunks them into a graph, generates and audits Q&A pairs, and embeds everything into Postgres + pgvector. That is the only place an LLM ever runs, and it runs on my machine, not in the request path.

Why I built it

I wanted a portfolio that answers as me, not as a chatbot guessing. So I built the retrieval layer myself: curated, grounded, reproducible, and cheap enough to run on the edge with zero per-message LLM cost. The constraints made it better - deterministic answers are easier to trust, easier to debug, and they don't surprise me in front of a recruiter.

Where it's at

Shipped and running. Published as a set of @cosimi/* packages on npm - an SDK plus adapters - where the infra drivers (Postgres, embedders, the offline LLM) are peer dependencies the consumer injects, so the core stays Workers-safe and adapter-agnostic. Deployed all-Cloudflare: Pages + Workers + Hyperdrive to Neon Postgres, embedding with Workers AI in prod and ollama in dev - one bge-m3 vector space across both.

Stack details

TypeScript, Postgres + pgvector, bge-m3 embeddings (ollama in dev, Workers AI in prod), Cloudflare Workers + Hyperdrive, Hono, Anthropic (offline ingest only), pnpm + Turbo monorepo, Vitest, Neon.