Answer cache

Cache Jev's answers, then run decide() again. Changing a threshold does not require another model call.

What makes a hit

The key combines the requested model, the question schema, and a canonical batch:

• Loki URL and tenant (demo fixtures have a separate identity).
• All source labels, including cluster and namespace.
• Sorted normalized patterns, with a count bucket for each pattern.
• Total and per-level count buckets, sampling omissions, and window length rounded to tenths of a minute.

Normalization removes timestamps, IDs, UUIDs, and long hexadecimal tokens. It preserves HTTP status codes, durations, resource measurements, and the full normalized line rather than just its prefix.

Count buckets are 0, 1, 2-9, 10-99, 100-999, and 1000+. Five occurrences and eight can reuse an answer; five and fifty cannot. Counts within a bucket are deliberately approximate. Use --no-cache when that approximation is inappropriate.

Editing a question's ID, type, instructions, or criteria causes a miss. Changing the canonical key format also invalidates old entries; old rows remain until cleared.

Model aliases

The key uses the model ID requested, not the version returned by Jev. An alias such as jev-latest can move without invalidating existing answers. Use a pinned version for repeatable results, or clear/disable caching when you need fresh answers from an alias. There is no TTL or automatic alias-change detection.

Storage and failures

The default file is ~/.cache/jev-logtriage/answers.sqlite3, or under $XDG_CACHE_HOME when set. Override it with --cache-db PATH.

SQLite stores hashes and answer JSON, not the Loki payload or the resulting decisions. WAL permits concurrent readers, with one writer at a time. seen_at updates on successful reads and writes.

An unavailable cache produces a warning and the run continues uncached. Malformed JSON rows are discarded and fetched again. Invalid Jev answers produce an analysis error rather than a success decision and are not cached.

• --no-cache: skip cache reads and writes.
• --cache-db PATH: use a different database.
• --cache-clear: delete cached answers before analysis.

Compare cached and uncached runs

scripts/bench-cache.sh

Requires TYPESAFE_API_KEY. The script uses its own temporary database and removes it afterward. It times the bundled demo with an empty cache, a warm cache, and caching disabled. Cold and uncached runs make API calls; warm should show hits and zero input tokens. These are single-run timings, not an accuracy benchmark. See the live verification notes for a sanitized-snapshot run.