mirror of
https://github.com/rohitg00/agentmemory.git
synced 2026-09-14 20:16:33 +08:00
efenex
e68d4ebe9b
* fix(summarize): chunk large sessions to fit LLM context window JSONL-imported sessions can have far more observations than the 500-cap MAX_OBS_PER_SESSION that constrains native sessions. mem::summarize previously built one prompt containing every observation and shipped it as a single LLM call, which exceeded the provider's context window for sessions >~7,000 observations and returned an unhelpful 400 from upstream — silently leaving large bulk-imported sessions out of the semantic tier. Approach: map-reduce inside mem::summarize. - Sessions ≤ SUMMARIZE_CHUNK_SIZE (default 400) take the legacy single-call path with no overhead - Larger sessions are split into chunks, each summarized with the existing per-session prompt in parallel batches of SUMMARIZE_CHUNK_CONCURRENCY (default 6), and partial summaries merged via a new REDUCE_SYSTEM prompt - Per-chunk retry-once on transient parse / provider errors - Persistently-failing chunks are skipped (not propagated) so a flaky chunk doesn't waste 30+ already-completed LLM calls on the same session - Bail with too_many_chunks_skipped only if >50% of chunks fail Companion operator tool: scripts/backfill-imported-sessions.sh walks jsonl-imported sessions and POSTs mem::summarize per session, with project / agent / obs-count filters, cost estimation, and per-failure payload dumping for debugging provider rejections. Validated locally against a real corpus: - 5,392-obs session (14 chunks, c=6): 39s - 10,704-obs session (27 chunks, c=6): 34s - 105,966-obs session (265 chunks, c=50): handler completes server-side and persists - 52-session bulk backfill → 25 new semantic facts + 6 new reflect insights produced by consolidate-pipeline Known limit: iii-engine has a hardcoded 180s function-invocation timeout. Sessions large enough that chunked summarize wallclock exceeds that will return a timeout/500 to the HTTP client even though the handler completes and persists server-side. High-RPM providers (Novita / DeepInfra / DeepSeek typically allow 100+ concurrent) can raise SUMMARIZE_CHUNK_CONCURRENCY to push the cliff well past any realistic session size. True fix is an async-job pattern; left as follow-up. - src/prompts/summary.ts: add REDUCE_SYSTEM + buildReducePrompt - src/functions/summarize.ts: chunking, retry, skip, parallelism - test/summarize.test.ts: 9 cases covering single-call path, chunking, env-override, retry-then-success, persistent skip, too-many-skipped bail, provider error after retry, concurrency - .env.example: document SUMMARIZE_CHUNK_SIZE / _CONCURRENCY - .gitignore: agentmemory-debug/, data-*/ (operator artefacts) - scripts/backfill-imported-sessions.sh: bulk-import backfill tool 9/9 new tests pass; existing tests untouched. * fix(summarize): address CodeRabbit review on #472 Four nits flagged by the automated reviewer, all worth fixing: - scripts/backfill: add curl --connect-timeout + --max-time profiles (META_CURL_OPTS vs WORK_CURL_OPTS). Metadata reads fail fast and retry on transient blips; LLM-backed work calls get a wide 30-min cap and no retry (retrying a half-finished LLM job double-spends). - scripts/backfill: sanitize sessionId before joining with DEBUG_DIR in dump_failure() (otherwise a session id containing `/` or `..` could escape the debug dir). UUIDs in practice, but the server doesn't enforce that. - scripts/backfill: switch the observations query to `--get --data-urlencode "sessionId=$id"` so special characters can't corrupt the query string. - scripts/backfill: guard `jq` on summarize + consolidate responses with `jq -e . </dev/null 2>&1` first. iii's HTTP layer occasionally returns non-JSON (HTML 5xx, empty body on timeout). Without the guard, `set -e` aborts the whole backfill loop on a single bad response — now it logs `invalid_json_response` and moves on. - test/summarize.test.ts: fix `vi.mock("./audit.js", ...)` path to `"../src/functions/audit.js"`. The old path resolved to `test/audit.js` (nonexistent), so the mock was a silent no-op. Tests passed anyway because `safeAudit` writes to a mocked KV. 9/9 tests still pass; backfill dry-run still resolves the corpus cleanly.