Kihagyás

vault-ko-ingest extraction-prompt tightening (vocab v3, 2026-05-19)#

Trigger#

Wave-1 Memgraph entity-cleanup analysis (../06-Audits/2026-05-19 Memgraph entity-cleanup analysis) found 8,975 sc≥1 KO-DB-rooted entities, of which 51% single-mention and 44% ≥3 token sentence-fragment. The two-tier vault-graph-diff Jaccard against the deterministic Tier-2 graphify extraction was 0.0070 — the LLM extraction was generating noise at scale (quoted strings, hex colors, code fragments, sentence fragments as "entities").

Changes#

Vocab version: 2026-05-17-v2-38pred → 2026-05-19-v3-38pred-antinoise7

7 new anti-noise rules#

  1. Quoted-string subject ban — subject cannot start with ", ', „. Extract the unquoted concept instead.
  2. Hex/URL/port/path/numeric → object-side — if the subject is purely #abc123, http://..., a port number, ./foo.py, or a literal numeric, move it to the object field of an annotation triplet.
  3. Code/operator-expression ban — def, class, function, return, import, lambda, and operator-laden expressions (x = y + 1, a += b) are never valid subjects.
  4. 60-char / 4-token subject cap — subject must be ≤60 chars AND ≤4 tokens. Multi-sentence fragments are forbidden.
  5. ALL_CAPS / snake_case-only ban — pure KGC_ADMIN and my_function_name are code-symbols, not concepts. Skip unless explicitly defined as a named entity.
  6. Fenced code-block exclusion — content inside … documents syntax, not domain concepts.
  7. Single-mention confidence floor ≤0.5 — if a triplet appears once in a single source-file with no prior mention, cap confidence at 0.5.

Plus a 5-case ### Anti-noise examples block with ✗ Wrong / ✓ Right contrasts.

Activation#

Default-on as of 2026-05-19 — every new vault-ko-ingest --file <path> run uses the v3 prompt. The vocab_version field in pending/<hash>-request.json lets downstream subagent-response handlers distinguish the eras.

Verification target#

vault-graph-diff Jaccard 0.0070 → ≥0.05 (Phase-4 acceptance gate). Phase-1+2 deletes alone won't get there (deletes only shrink the denominator); selective re-extract of the 5,524 sentence-fragment source-files with the v3 prompt is needed — scheduled 2026-05-22-23. See ../06-Audits/2026-05-19 Memgraph cleanup Phase-3 next-step plan.

Backup#

Pre-tightening prompt-template snapshot: .vault-ko/prompts/vault-ko-ingest.py.bak.20260519-pre-tighten (10,575 bytes).