# klimatsearch

> Hybrid search and MCP server for Boverket Klimatdatabas. One Go binary: SQLite FTS5 + sqlite-vec, REST, MCP, operator console at GET /. Self-host is ungated (no Unkey, no MPP). Cite Boverket Klimatdatabas. Not affiliated with Boverket.

The running process also serves an operator console at `GET /` (`html/template`; Query, Data, Ingest, Webhooks, Lab). Catalog `dkbr` is Danish BR; DatasetVersion is BR18/BR25. Agents should still call MCP `/mcp` or REST `/api`, not scrape the HTML. JSON Resource upsert: `POST /admin/resources`. Webhooks: `GET /webhooks` or `KLIMAT_WEBHOOK_URL` (Slack text, Discord content, else JSON). Events: `catalog.changed`, `catalog.unreachable`, `ingest.failed`.

An agent setting up a client or a self-host instance should load [docs/AGENT-SELFHOST.md](https://github.com/mong-x/klimatsearch/blob/main/docs/AGENT-SELFHOST.md) first. Search with the cheapest path that can hit the labeled Resource: FTS, then hybrid (`vector=true`), then rerank (`rerank=true`) only for sibling grades. REST defaults vector and rerank off. MCP uses whatever the process loaded (`KLIMAT_RERANKER=onnx` means MCP always reranks). Scale by replica-local SQLite plus HPA; do not share one SQLite WAL across writers. Do not load zerank-2.

## Docs

- [Agent self-host playbook](https://github.com/mong-x/klimatsearch/blob/main/docs/AGENT-SELFHOST.md): Cost, MCP/REST, load, and HPA for a self-hosted instance
- [README](https://github.com/mong-x/klimatsearch/blob/main/README.md): Product, REST, MCP, retrieval quality
- [SELFHOST.md](https://github.com/mong-x/klimatsearch/blob/main/docs/SELFHOST.md): F2LLM + BGE/zerank files, compose, check-models
- [LAUNCH.md](https://github.com/mong-x/klimatsearch/blob/main/docs/LAUNCH.md): Human steps (weights, Unkey/MPP hosted lane, k8s)
- [CONTEXT.md](https://github.com/mong-x/klimatsearch/blob/main/CONTEXT.md): Glossary (Catalog, Resource, Hit, Guard)
- [OpenAPI](https://mong-x.github.io/klimatsearch/swagger/): REST spec
- [Golden queries](https://github.com/mong-x/klimatsearch/blob/main/testdata/golden-queries.json): 44 labeled sibling/exact/cross-lingual queries

## Deploy

- [compose.onnx.yml](https://github.com/mong-x/klimatsearch/blob/main/deploy/compose.onnx.yml): Docker Compose with F2LLM + BGE reranker
- [deployment.onnx.yaml](https://github.com/mong-x/klimatsearch/blob/main/deploy/k8s/deployment.onnx.yaml): Kubernetes Deployment (4 GiB, reranker on)
- [hpa.onnx.yaml](https://github.com/mong-x/klimatsearch/blob/main/deploy/k8s/hpa.onnx.yaml): HorizontalPodAutoscaler for CPU/memory

## Optional

- [GitHub Pages](https://mong-x.github.io/klimatsearch/)
- [Boverket Klimatdatabas](https://www.boverket.se/sv/klimatdeklaration/klimatdatabas/)
- [Boverket product sheet pattern](https://klimatdatabasen.boverket.se): origin URL is `/detaljer/{category_code}/{resource_id}`