Chunk and index files are disposable runtime state. Incremental prefixing skips
records whose chunk and page hashes still match. A complete scan removes
surplus records for deleted pages, and the prefixer invalidates the BM25 index
before changing its chunk set so a mixed stale index is not served.
Prefix and BM25 build operations share the vault-wide mutation lock with every
other writer; a busy vault fails closed instead of publishing a partial index.
Contextual-prefix privacy
Synthetic prefixes use only local frontmatter and page text. The Anthropic API
and claude subprocess tiers can send page bodies off-machine and therefore
require the user's explicit consent plus --allow-egress. Never infer consent
from an API key or installed binary. Preview the scope first and state which
provider will receive what data.
Remote Ollama endpoints also require explicit approval and
--allow-remote-ollama; the default reranker accepts localhost only.
Query
For a strictly read-only lookup, use the prebuilt BM25 index:
For an explicitly requested rerank, omit --no-rerank. The default is Ollama's
multilingual nomic-embed-text-v2-moe model (approximately 958 MB); the product
never pulls it automatically. To use an already-installed, smaller,
English-oriented v1.5 model, pass --model nomic-embed-text explicitly.
Nomic models use search_query: for the query and search_document: for
candidate text. Nomic v2 has a 512-token input context and Ollama truncates
longer embedding inputs by default; BM25 still scores the complete chunk.
Embeddings are cached by exact model, input scheme, and hash of the exact
prefixed input. A missing local Ollama service, missing selected
model, unusable vector, or any candidate embedding failure falls back for the
complete result set to the original BM25 order; it never mixes cosine and BM25
score scales.
Query input is bounded at 8,000 normalized characters and result counts must be
between 1 and 1,000. Oversized queries and invalid limits fail with an
actionable usage error instead of looking like an empty successful search.
An untagged model request matches only the installed untagged name or its
:latest alias; select any other tag explicitly.
Accept only relative chunk and page paths whose resolved targets remain under
$VAULT/.vault-meta/chunks/ and $VAULT/wiki/ respectively.
Reject hashless legacy chunk records and require chunk-body, page, and index
hashes to match before a cached record can be built or served.
Reject absolute paths, symlink escapes, missing pages, mismatched chunk IDs,
changed page hashes, and stale index/chunk hash pairs.
Rerank the full candidate set, then deduplicate by page, then apply --top.
An empty index is an honest no-result state. A missing or corrupt index makes
retrieve.py exit 10 with a stable rebuild command; callers fall back to the
standard vault query/text-search path and do not fabricate matches.
Do not cite benchmark percentages unless a reproducible vault-specific
benchmark produced them.
Checkpoint
Observe cache readiness and privacy boundaries, think about whether lexical or
semantic ranking is needed, verify returned paths and source freshness, and
grow by measuring retrieval misses against a maintained local query set.
