8bu.dev

Personal portfolio with a chat that answers questions about me — deterministic GraphRAG, no LLM at query time.

---

A TanStack Start web app plus a Hono chat backend. Ask it something; it embeds the query, retrieves ranked Q&A pairs and document chunks from a curated corpus, and streams the top answerable hit over SSE. No generation, no hallucination — every reply traces to a source pair.

Cosimi

Retrieval runs on cosimi — my GraphRAG SDK (@cosimi/*). The chat embeds a query, walks pgvector-indexed pairs/chunks, and returns the best-grounded match deterministically. createCosimi({ sql, embedder }).retrieve() is the whole query path; a canonical override + seed-pair layer sit in front for curated, on-topic answers. Ingest (LLM pair-generation, embedding, upsert) is an offline operator step, never run at request time.

Layout

apps/web/        TanStack Start portfolio — chat UI + MDX artifacts + OG gen
services/api/    Hono chat backend — Node (index.ts) + Workers (worker.ts) entries
services/admin-api/  ingest backend — pair-generation, embedding, upsert
packages/core/   chat DTOs, valibot env, normalize(), request-scoped sql
packages/db/     postgres adapter, migrations, migrate CLI, seed
packages/logger/ pino + redactInput, Workers-safe
corpus/*.md      long-form docs → documents/chunks/pairs
seeds/           curated Q&A pairs (YAML)

Tech

Node 22 · pnpm 11 · Turbo 2 · TS 5.7 · Hono · postgres.js · Postgres 16 + pgvector · valibot · pino · oxlint/oxfmt · vitest · Vite 7 · TanStack Start/Router · Tailwind v4 · Cloudflare Pages + Workers + Hyperdrive + Neon.

Develop

pnpm install
pnpm dev     # Docker → PG (:5434) → migrate → seed → web (:5174) + api (:3010)

The web dev server proxies /api/* → localhost:3010, mirroring the prod Worker route 8bu.dev/api/*.

Gates

pnpm -r typecheck
pnpm lint
pnpm format:check
pnpm -r --workspace-concurrency=1 test   # DB tests race when parallel

See CLAUDE.md for the full architecture, answer cascade, and deploy flow.