codex mechanicus search (spec-014): ground warhammer lore in binaric.tech via codex_search tool

2026-07-13 20:50:08 +02:00
parent a514ff652c
commit 09871b9b95
6 changed files with 391 additions and 1 deletions
@@ -0,0 +1,59 @@
+# SPEC-014 — Codex Mechanicus search
+
+Luma is an Adeptus Mechanicus tech-priest; her lore has a real home —
+the priest's own Codex Mechanicus at `binaric.tech` (an Astro/MDX
+archive, five tongues). A `codex_search` function tool lets her consult
+that archive and answer from sourced inscriptions instead of inventing
+lore. The index is public but still untrusted by the time it reaches a
+prompt: the fetch is SSRF-guarded (SPEC-011 shares `guard_url`),
+size-bounded, and every returned field is sanitized (SAF-03). Luma-only;
+active only when `enable-codex = true`.
+
+### CDX-01 — codex_search is offered as a tool (coverage: test)
+
+When `enable-codex` is true, the chat call's `tools` list includes a
+`codex_search` function (`query` string, optional `lang`) next to any
+IGDB / fetch_url tools. When false, it is absent.
+
+### CDX-02 — The index is fetched safely and cached (coverage: test)
+
+The index URL (`codex-index-url`, default
+`https://binaric.tech/search-index.json`) passes the SSRF guard before
+any network call, is read under a byte cap (`codex-max-bytes`, default
+4 MB) with a download timeout, and is cached in memory for
+`codex-cache-ttl` (default 3600 s) so repeated searches do not re-fetch.
+
+### CDX-03 — Ranking weights title over summary over body (coverage: test)
+
+The query is tokenized (stopwords dropped); each inscription is scored
+by term hits weighted title (8) > summary (3) > body (1). Results are
+returned highest-score first, each as `{title, summary, collection,
+url}`, with `url` absolute against the site origin.
+
+### CDX-04 — Language is preferred, with fallback (coverage: test)
+
+Results are filtered to the requested `lang` (en, de, eo, no, uk;
+default en; unknown codes fall back to en) by the language segment in
+each inscription URL. If no inscription in that tongue matches, the
+search falls back to all tongues rather than returning nothing.
+
+### CDX-05 — Results are sanitized and failure is reported (coverage: test)
+
+Each `title` and `summary` is passed through `sanitize_external_text`
+and length-capped (`codex-summary-chars`, default 500). An index that
+cannot be fetched or parsed returns an `{error: ...}` dict the model can
+relay — `search` never raises.
+
+### CDX-06 — Searches are metered per user (coverage: test)
+
+Each `codex_search` increments a per-user daily counter; over
+`codex-daily-per-user` (default 50) the tool refuses with an error
+result without touching the index. The budget gate (SAF-04) still
+applies to the surrounding model calls.
+
+### CDX-07 — Luma cites the codex, not invention (coverage: manual)
+
+With the persona grounding line, when a pilgrim asks Cult Mechanicus
+lore Luma consults `codex_search` and answers from it, offering the
+`binaric.tech` link to read the full inscription rather than
+hallucinating. Verified live on ggg.