Electric P8 2027

Part of eplan-cloudflare-rags. The local action server lives in eplan-mcp.

Remote MCP server that lets Claude search the EPLAN 2027 API documentation through full-text/keyword search (SQLite FTS5 + bm25), over the wiki produced by eplan_wiki_scraper.py (classes/interfaces/structs/enums bundled per type with every member inlined — see that script’s docstring for why).

Companion to cloudflare-rag-eplan-p8/, not a replacement. That one does semantic search (Vectorize + bge-base) over the 2026 docs. This one does keyword/exact-match search (FTS5) over the 2027 docs. Measured against each other on real queries: FTS5 wins for “what’s the signature of X” / “what does property Y do” (most API-reference lookups); semantic search wins when the query uses none of the source’s vocabulary at all. Use both — they fail differently.

Architecture

Claude Code  -->  MCP (Streamable HTTP)  -->  Cloudflare Worker (eplan-wiki-2027)
                                                   |
                                              D1 (SQLite + FTS5, bm25 ranking)
                                              1,754 documents / ~36 MB

No embeddings, no Vectorize, no Workers AI call per query — a docs_fts MATCH query against D1 is the entire search path.

Install the MCP in Claude Code

claude mcp add eplan-wiki-2027 -- cmd /c npx mcp-remote https://rag2027.covaga.xyz/mcp
claude mcp list   # should list "eplan-wiki-2027"

Available Tools

Tool	Description
`eplan2027_search`	Keyword/full-text search. Best for exact or near-exact names.
`eplan2027_get`	Fetch one file’s full content by path (from a search result).
`eplan2027_stats`	Document count / index info.

REST Endpoints

Base URL: https://rag2027.covaga.xyz.

Method	Endpoint	Description
GET	`/health`	Health check
POST	`/search`	Keyword search (body: `{"query": "...", "topK": 8, "kind": "bundle"\|"standalone"}`)
GET	`/file?path=...`	Full content of one file by path
GET	`/stats`	Document count

curl -X POST https://rag2027.covaga.xyz/search -H "Content-Type: application/json" \
     -d '{"query": "FindAction", "topK": 5}'

Re-ingesting after a wiki re-scrape

The D1 schema (schema.sql) defines docs + an external-content FTS5 table (docs_fts, synced via triggers — every INSERT INTO docs auto-populates it).

# 1. Apply schema (idempotent — DROP+CREATE)
cd worker && npx wrangler d1 execute eplan-wiki-2027 --remote --file=../schema.sql

# 2. Re-add the temporary bulk-load endpoint's secret (see worker/src/index.ts,
#    POST /admin/ingest -- bound parameters, avoids the SQLITE_TOOBIG limit
#    `wrangler d1 execute --file` hits on large inlined SQL string literals)
npx wrangler secret put INGEST_TOKEN   # paste a random token
npx wrangler deploy

# 3. Ingest
cd .. && python ingest.py --url https://rag2027.covaga.xyz --token <the token>

# 4. Remove the secret again -- the endpoint no-ops (401) without it
cd worker && npx wrangler secret delete INGEST_TOKEN

Deployment notes

D1 database: eplan-wiki-2027 (632a6de2-d0f2-498c-921e-cb1e77b040aa, region WEUR).
Custom domain rag2027.covaga.xyz was attached via the account-level Workers Custom Domains API (PUT /accounts/{id}/workers/domains) rather than wrangler deploy’s automatic route setup — the deploying API token had #d1:edit/#worker:edit but not the zone-scoped Workers Routes permission wrangler tries first. If re-deploying hits the same A request to the Cloudflare API (/zones/.../workers/routes) failed error, either add that zone permission to the token or call the Custom Domains endpoint directly instead of relying on the routes block in wrangler.jsonc.
This machine sits behind a Zscaler TLS-inspecting proxy; Node’s own CA bundle doesn’t include Zscaler’s root, so wrangler/npm need NODE_EXTRA_CA_CERTS pointed at an exported copy of it or every HTTPS call fails with UNABLE_TO_GET_ISSUER_CERT_LOCALLY (surfaces from wrangler as a generic fetch failed, including during wrangler login’s OAuth code exchange).

This page is synced from Canary-ECAD/eplan-cloudflare-rags.