Skip to content

v0.3.7

Choose a tag to compare

@tonydzi tonydzi released this 06 Nov 00:38
9bde194

Recall: the cross-encoder rerank is off by default. sgm-ask, the MCP tool and eval/run_eval.py now order candidates by dense e5 similarity; --rerank or SGM_RERANK=1 brings the cross-encoder back. Measured before the change: on our 184-question gold set the reranker cut body nDCG@12 from 0.815 to 0.640, and on DolphinBench data (Riley, 50 tests, 1000 sessions) Recall@5 fell from 0.61 to 0.48; it still helped short title-style queries. Dense recall is also faster on CPU (about 0.04 s against 2.8 s per query on that slice). rerank_candidates() takes sims= and keeps dense order when no cross-encoder is passed.

uvx install lines work on a newer default Python. On an Intel Mac whose uv defaults to Python
3.14, uvx --from 'sqlite-graph-memory[lite]' ... fell back to release 0.3.5, which has no [lite], and
sgm-index stopped with a hint to install [embeddings]. Apple Silicon below macOS 14 resolves the same
way. The uvx lines in the docs now pin --python 3.12; server.json does too, and that reaches the MCP
Registry with this release.

Shorter demo gif. It now shows the top 5 of each list, so the graph hit fits on one screen.

sgm-index: a rerun re-embeds only what changed. A manifest keyed on the content hash and the index settings lets sgm-index skip untouched notes; size and mtime are the fast path, the hash decides the changed set, and a corrupt manifest falls back to a full build. The default encoder still comes from load_encoder(), so [lite] installs keep working without torch. Thanks @Jason-Vaughan (#85), who picked up #17.

Recall: folder-style wikilinks resolve in the graph hop. [[folder/note]] links now reach their note in expand_1hop the same way sgm-stats counts them, preferring the exact vault-relative path; the vault root is computed once per index, not per link. Thanks @Jason-Vaughan (#84), who picked up #71.

Bench: tangleclaw-style reference row. A stdlib-only pipeline (BM25 + FTS5 + 1-hop wikilinks + superseded_by lineage) in the style of TangleClaw's markdown continuity, reproduced in CI. Same R@5 as bm25; its gain is on superseded decisions and abstention, with no domain words in the abstention gate. Thanks @Jason-Vaughan (#87), who picked up #21.

eval/run_ablation.py: one run sweeps the retrieval knobs. Scores a grid of cells (hops, ghops, gmax, topk, entity gate) on one gold file, reusing the same pipeline functions as sgm-ask and the metrics from run_eval.py; the dense pass runs once per question and each cell slices the widest pool, so only the rerank repeats. The default grid changes one axis at a time against 1-hop; --grid takes multi-axis cells. Thanks @Jason-Vaughan (#86), who picked up #19.

Light backend: a fallback to torch is no longer silent. When the ONNX models fail to load and recall falls back to torch, the answer now starts with a NOTE: line naming the reason and pointing at sgm-doctor. Through MCP the agent sees only the answer, so the stderr line alone hid the slow path (#89).

Docs: install commands use [lite]. Now that v0.3.6 is on PyPI with the [lite] extra, README, llms.txt and the example MCP config install the torch-free stack; [embeddings] stays documented for a GPU.