[jb-thery]

Problem

Ragmir 6.4.0 still answered too many questions its sources cannot answer when no embedding model
is configured, and its ranking left room on real repositories:

• with local-hash, only 12 of 90 off-topic and 1 of 110 near-domain questions were refused on
  three real bases;
• a passage that only the keyword search returned could stand as an answer on one shared everyday
  word, and rare function words such as "what" or "de" could rank passages on their own;
• on a small corpus whose headings share one word, a heading boost could bury the passage both
  retrieval channels ranked first, and the search then returned nothing;
• rgr evaluate timed concurrent cases, so its latency read several times higher than real searches;
• the landing hero showed the GitHub and npm buttons in the left column instead of under the
  terminal demo.

Changes

• Ranking: function words weigh 0.3 in the keyword query; passages whose file name or heading
  path holds words of the question gain up to 20%, ignoring words that more than half of the
  passages hold. Ranking policy version 9.
• Abstention: with a semantic model, a keyword-only passage needs half of the question's IDF
  weight. With local-hash, search answers nothing when none of the first five passages holds 30%
  of the weighted words, or when words found nowhere in the index carry half of the question's
  weight (a missing word weighs as if one passage held it). Coverage credits fuzzy matches and
  identifier parts.
• Speed: query-independent work is memoized per passage text and path.
• GPU ingestion: ingestEmbeddingDevice: "webgpu" embeds document batches on the GPU and falls
  back to the CPU; queries stay on the CPU.
• Diagnostics: rgr doctor and rgr upgrade name an explicit rgr ingest when external
  extractors block automatic repairs; rgr evaluate takes latency from sequential warm searches.
• Landing: the GitHub and npm buttons sit under the terminal demo, aligned right, and the hero
  proof shows the new measurements.
• Docs: both READMEs, configuration, migration, CLI, troubleshooting, llms.txt, and the
  benchmark guide.
• CI: the macOS portability job moves from macos-14, which GitHub retires on November 2,
  2026 and already fails during scheduled brownouts, to macos-15.

Measured results

Ragmir 6.4.0 against this branch on the same six indexes (three real bases, each with its semantic
model and with local-hash), two rounds, warm searches on an Apple M3 Max:

Measure	6.4.0	This branch
Answer first, Java/TypeScript monorepo (e5-base)	44%	54%
Answer first, same monorepo with local-hash	28%	35%
Off-topic questions refused	66 of 180	133 of 180
Near-domain questions refused	24 of 200	53 of 200
Correct answers refused	0 of 500	0 of 500
Median search, monorepo	39 ms	36 ms

The answer ranks first more often on all six indexes. One loss: local-hash on this repository
finds three fewer answers in the top 5. Ideas tested and rejected on the same indexes: file-level
rank fusion, joined query words, a test-file prior, score-based fusion, and the file path in the
embedded text.

Verification

• pnpm validate: Biome, Knip, 570 Core and 27 landing tests with coverage gates, API, CLI, MCP,
  and public-surface smoke checks, builds, publint, and release checks
• an absent-word stress test on 250 answerable questions: the semantic rule stays at its 6.4
  behavior, and local-hash refuses 2 more of the altered questions
• pnpm bench:competitors: Ragmir reaches the top 3 for 83 to 88% of questions, against 63% for
  LangChain.js and 79% for MiniSearch
• the hero checked in a browser at 1440 px and 375 px in English and French