You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The engine now embeds each prompt and ranks the workspace's memories, rules, and
skills, so the hook delivers only the few most relevant full bodies per turn with
a graceful lexical fallback. Cross-lingual matching leans on embeddings; the
local backend stays cloud-free behind the KnowledgeBackend seam.
Signed-off-by: Sertan Helvacı <sertanhelvaci@icloud.com>
|`PATHRULE_EMBEDDING_PROVIDER` = `voyage`\|`openai` + `PATHRULE_EMBEDDING_API_KEY`| Semantic memory search: embeddings computed on write, cosine-ranked at query time, stored locally| lexical + path-scoped retrieval |
139
+
|`PATHRULE_EMBEDDING_PROVIDER` = `voyage`\|`openai` + `PATHRULE_EMBEDDING_API_KEY`| Semantic relevance, two places: (1) `get_context` memory search, and (2) the hooks — each prompt is embedded and the most relevant memory/skill bodies are ranked and injected just-in-time. Embeddings are computed on write and stored locally; only the prompt is embedded at runtime. | lexical + path-scoped retrieval (hooks still inject, ranked by keyword overlap instead of meaning)|
140
140
141
141
Both degrade gracefully. On timeout or a missing key, Pathrule falls back to the deterministic
142
-
path, the same fallback discipline the cloud edition uses.
142
+
path, the same fallback discipline the cloud edition uses. The embedding key is what turns the
143
+
hooks from "inject the path's titles" into "inject the few bodies this prompt actually needs," so
144
+
it is the highest-leverage key for context quality and token cost.
0 commit comments