Tech stack

Languages, frameworks, and the LLM backend configuration behind Vessios.

Backend services

• wiki-worker — Python, FastAPI. Owns all domain and business logic. Dependencies are pinned exactly (==, not ^/>=) to avoid dependency-confusion risk.
• wiki-gateway — Rust, Axum. A thin, type-unaware passthrough reverse proxy; it forwards request bodies without parsing them, so API type generation targets wiki-front, not the gateway. unsafe Rust is disallowed project-wide.
• wiki-front — Next.js, TypeScript. Consumes the API described in the API reference.

Infrastructure

• Firestore — tenant, page, source, and knowledge-graph data, with per-tenant subcollection isolation.
• Vertex AI — embeddings and Vector Search; also hosts Claude when routed through Vertex.
• Redis — caching, rate limiting, and the semantic query cache.
• Pub/Sub — asynchronous pipelines (file ingest, export).
• BigQuery — a three-tier logging system (query_logs, usage logs, audit logs).
• Cloud Run — hosts both wiki-worker and wiki-gateway; deployed via GitHub Actions, with infrastructure managed as code through OpenTofu across dev/staging/prod environments.

Embeddings

Production uses text-embedding-005 (Vertex AI, 768 dimensions, Japanese-capable). Local development uses multilingual-e5-small (384 dimensions). Because the dimensionality differs, local and production vector indexes are not interchangeable.

LLM backend selection

wiki-worker selects its LLM client through a factory, controlled by the LLM_BACKEND environment variable. Container wiring and api/main.py depend only on the factory — never on a concrete client — keeping the dependency direction inward per Clean Architecture.

Testing

• wiki-worker: pytest, with all external services mocked by default (no live API calls or charges in CI).
• wiki-gateway: cargo test plus cargo clippy --all-targets -- -D warnings (zero warnings tolerated).
• Evaluation harness (wiki-worker/eval/): offline and mock-backed by default; a --live flag opts into real Claude calls for golden-query regression and mode A/B testing.