Repository navigation
feat(s032): complete slice 032 (semantic memory and hybrid retrieval) - #45
Merged
Merged
Conversation
… store measured, the G1 plan, and the slice opens ROADMAP row 032 to in_progress. NOTES.md records the decisions (brute force and pgvector, no sqlite_vec; local first, hosted only by configuration and never a fallback; the tier off when local cannot run; the embedder recorded per vector; no text to a hosted embed API without opt-in), the measurements (local 86 ms and 0.858/0.062 on the thresholds, 463 MB of XLA, no Windows target; hosted 208 ms and 0.71/0.752; pure Elixir 23 s), the deps, the eight-line plan, the two-item manual queue, four deviations. mix.exs and mix.lock carry nx, exla, bumblebee and pgvector; scripts/embed_bench.exs is the measurement. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…d retriever, recall tool, memory page tab The embedder behaviour with the local Bumblebee serving (explicit download, never a hosted fallback), the hosted alternative behind configuration, and the fake for the suite; the vector on the memories row with the model that produced it, searched brute force on SQLite and through pgvector on Postgres (Postgrex types defined for that build); the observer after each turn under the memory task supervisor with the 0.92 dedupe; the retriever fusing vector and full-text hits by reciprocal rank with recency decay and a cosine floor; the recall block in the volatile tier under its own cap with the 030 receipt; the recall tool; the memory page's Semantic tab with status, search, provenance, delete, pin and the download action. VERSIONS rows for nx, exla, bumblebee, pgvector, hnswlib; sqlite_vec recorded as not used. docs/01, 02, 05 amended. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…y, and the tier's status; exla declared on Linux and macOS hosts only AC7's fallback clause: no loadable extension, so the packaged binary proves the vector store in force with three fake vectors rolled back (binding, exit 4), and records whether the EXLA NIF loaded and what the semantic tier's status is in the bundle (informative). xla ships no Windows archive, so a Windows host does not declare exla and the tier is off there by design. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…oses the assistant row Found by slice 032's AC6 run on nvidia:nemotron: the model streamed one newline and then its tool calls; validate_required counts whitespace as blank and the row was refused. Signed-off-by: Ayla Croft <aylacroft@proton.me>
References the red test in the previous commit: String.trim before the (no text) substitution, so a turn that is only tool calls after a newline keeps its assistant row. Signed-off-by: Ayla Croft <aylacroft@proton.me>
The turn after a tool result raised invalid tool_call inside req_llm 1.24.0 on the first live run of slice 032's recall; context/1 made public (@doc false) so the red can name it. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…and arguments References the red test in the previous commit. Every live turn after a tool result had failed with invalid tool_call since 011; the fake provider builds no context, so only a live run could see it. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…er the tree, EXLA scoring in the brute store Package run 35600216451: on Linux the EXLA NIF does not load in Burrito's musl ERTS (__libc_single_threaded: symbol not found) and exla's application start took the release down; the release now loads exla without starting it and Bumblebee.exla/0 starts it on demand, remembering a failure as the tier's reason. On macOS the NIF loaded (EXLA=ok). The vector check ran before the Repo on every OS; it is now a probe child before the endpoint. Brute.search scores 10,000 rows as one EXLA product (72 ms p50, was 531 ms in Elixir); the Elixir path stays for machines without the backend, held to the same numbers by a test. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…shots Signed-off-by: Ayla Croft <aylacroft@proton.me>
…d where the model is on disk (local_model tag) Signed-off-by: Ayla Croft <aylacroft@proton.me>
Signed-off-by: Ayla Croft <aylacroft@proton.me>
The fake provider's failures and objects no longer outlive the test that set them (llm_test and SessionCase clear on exit), and the test helper waits for the boot receipt's Task before the sandbox goes manual, so the boot-receipt tests no longer depend on a runner's speed. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…in :load mode Mix refuses exla: :load while trinity's applications list depends on it (package run 35603384655). Out of the applications list, exla is on the code path, started on demand by Bumblebee.exla/0 in every environment, and the release loads it without starting it. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…w it is carried Signed-off-by: Ayla Croft <aylacroft@proton.me>
…d query is never short The postgres job returned two of three rows (run 35604770525): ef_search's 40 candidates were other entries and dead tuples, and the WHERE clause came after. SET LOCAL hnsw.iterative_scan = strict_order inside the search's transaction. Signed-off-by: Ayla Croft <aylacroft@proton.me>
… store does Signed-off-by: Ayla Croft <aylacroft@proton.me>
…ef; package sizes in docs/perf.md Signed-off-by: Ayla Croft <aylacroft@proton.me>
…again Alarm's doc says clear/0 runs once a signer signs again; nothing calls it. The fips leg's receipts test met an alarm an earlier test left set (run 35606884798). Signed-off-by: Ayla Croft <aylacroft@proton.me>
References the red test in the previous commit: sign_if_needed/2 and sign_checkpoint/1 clear the alarm on a successful signature when it is set, which Alarm's doc has promised since 024. Signed-off-by: Ayla Croft <aylacroft@proton.me>
PROOF.md with every criterion's evidence, the coverage row (79.01 at 732238c), NOTES findings 12 and 13, and the ROADMAP row to done. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…has no argument to mistake for a file Package run 35608084951, macOS: 'No file named --smoke', exit 1, between the fourth line and the fifth. The workflow sets the variable; --smoke still works where the race is won. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…tests Every turn now runs the retriever, and the postgres job's 100-session and 20-writer tests refused checkouts on the sandbox's one shared connection (runs 35610164189 and 35610171809): queue_target 5 s and queue_interval 30 s on the test pools. NOTES findings 14 and 15. Signed-off-by: Ayla Croft <aylacroft@proton.me>
…erage row at 193800e Signed-off-by: Ayla Croft <aylacroft@proton.me>
… as a follow-up Signed-off-by: Ayla Croft <aylacroft@proton.me>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Slice 032: local embeddings (all-MiniLM-L6-v2 through Bumblebee and EXLA; an explicit download, never a hosted fallback: the owner's G0 decisions), the vector on the memories row with the embedder that produced it, brute-force search on SQLite (an EXLA product when the NIF is loaded) and pgvector's HNSW on Postgres, the observer that fills the semantic tier after each turn with a 0.92 dedupe, the retriever (reciprocal rank fusion of vector and full-text hits, recency decay, a cosine floor), the "Relevant memories" block in the volatile tier under its own cap, the
recalltool, and the memory page's Semantic tab with the download action.Evidence:
slices/032-semantic-memory/PROOF.md(AC1 to AC8; the AC6 GIF underproof/),NOTES.md(thirteen findings, four follow-ups),docs/perf.md(new).Outside the slice's scope, each red-then-fix because this slice's runs met them:
fix(s012)a whitespace-only assistant row was lost;fix(s011)req_llm's tool-call encoding failed every live turn after a tool call;fix(s024)the signer alarm never cleared. Two test flakes closed at their source.The packaged binary: the vector search passes in all three bundles; EXLA's NIF loads in the macOS bundle and not in the Linux one (Burrito's musl ERTS), so exla is started on demand and the Linux bundle boots with the tier off and says why (NOTES finding 2, a follow-up for a Linux backend); a Windows host does not declare exla.
Gate on the final commit: run 35607474111, all four legs green.
Manual queue for G4: AC2 (the real embedder's numbers;
TRINITY_LOCAL_MODEL_CACHE=<cache> mix test --only local_model) and AC6 (proof/ac6-recall.gif).Signed-off-by: Ayla Croft aylacroft@proton.me