[milliondreams]

Summary

Phases 3 and 2b of the kniv v7 NLP proposal (docs/proposals/kniv_v7_nlp_extraction_2026_11_07.md, rev. 7), on top of phases 1–2 (#327).

Phase 3: entity resolution with Locy

• A tested Locy program on the NLP feature page links each extracted :Mention to an existing :Entity. It combines weighted evidence (exact name, surname, embedding similarity) with MNOR, keeps the best match per mention, and creates one entity per new name, so variants of one name in a batch share an entity.
• Claims are stored against Mentions, so claims reach their entities through resolution. Speakers, forks (review with diff, then promote) and 300 × 300 scale are covered by tests.
• Testing that path exposed three engine bugs, each fixed here:
  • Fork writes to an auto-embedded label failed with "runtime not configured": a fork's writer now gets the Xervo runtime.
  • Promote duplicated auto-embedded vertices and skipped edges to them: the writer now stores auto-embedded values as Value::Vector, so the content UID matches the stored row.
  • DERIVE timed out on large rule sets: it re-joined its clause against the full cartesian product. It now narrows the MATCH to the derived pairs and indexes facts by entity (from over 30 s down to 54 ms at 300 × 300).

Phase 2b: word and sentence analysis

• uni.nlp.words: one row per word, with Universal Dependencies part of speech, morphology, lemma, and dependency head and relation. Indices count across the whole document, so head joins to another row's word.
• uni.nlp.sentences: one row per sentence, with sentiment, dialogue acts above the threshold, and every act's probability.
• Input is the same as for extraction, including conversations. options takes only alias, and a v1 model is refused.
• Rust and Python get analyze_words / analyze_sentences on db.xervo().

Known limits (filed)

• Locy: a similarity rule over a cartesian MATCH reads every pair, so entity resolution is quadratic #330: a Locy similarity rule over a cartesian MATCH reads every pair, so resolution is quadratic in mentions × entities.
• NLP: ship entity resolution as a built-in Locy module instead of a program copied from the docs #328: the resolution program has to be copied from the docs instead of shipping as a built-in module.
• Locy: no way to run a Locy program from Cypher, so a trigger cannot resolve entities after extraction #329: there's no way to run Locy from Cypher, so a trigger can't resolve entities after extraction.

Tests

• New integration tests: nlp_resolve (10 pass, 1 ignored) and nlp_analyze (8 pass against a mock model). The real-model tests are ignored and run in the nightly sweep. I ran them locally against kniv-v7-xlarge-u-en, including the feature page's query examples, and they pass.
• Locy TCK 558/558; uni-query lib 545; Python NLP tests 62.
• Clippy (as CI runs it), cargo fmt, ruff, stubtest, the strict docs build and the docs snippets all pass.
• Full uni-db integration suite under cargo test: 3,948 passed, 10 failed. The failures are scratch-directory, bulk-load and deadline tests. Each passed when rerun on its own: plain cargo test runs every test in one process, while CI's nextest gives each test its own.

🤖 Generated with Claude Code