Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(integrations): the shipped `answer_confidence` the wrappers gate on is not calibrated
Adeshina Lasisi2c6a2243ef70
diff --git a/docs/crewai.md b/docs/crewai.mdindex a7fa74ad..48016240 100644--- a/docs/crewai.md+++ b/docs/crewai.md@@ -101,13 +101,21 @@ except LayaTaskGuardError as e:----## 3. Calibrated Confidence Gating+## 3. Confidence Gating-`LayaCrewRouter` gates on calibrated `answer_confidence` (`max(p)`):+`LayaCrewRouter` gates on `answer_confidence` (`max(p)`):- **Automatic Fallback:** Specify `fallback_agent_index` to route ambiguous tasks to a human supervisor or general lead agent.- **Strict Guarding:** Set `raise_on_low_confidence=True` to raise `LayaLowConfidenceError` when a task cannot be matched to an agent role with sufficient confidence.+`answer_confidence` is the quantity temperature scaling fits and the one every calibration figure+in the repository is computed on, which is why the threshold reads it rather than the entropy+`confidence`. It is not calibrated as shipped: `laya-multilingual` ships no fitted temperatures at+all (`temperature: [1.0, 1.0, 1.0]`, empty `temperature_by_options`), and english's `choice:11+` is+refused by the loader's own clamp with a warning that says to treat it as uncalibrated. Fit and+validate any threshold on held-out data at the option counts your workload uses -- see the+README's [Calibration](https://github.com/NandhaKishorM/laya#calibration) section.+---## 4. Remote HTTP Server Deploymentsdiff --git a/docs/langchain.md b/docs/langchain.mdindex a572ca37..5660c6ab 100644--- a/docs/langchain.md+++ b/docs/langchain.md@@ -74,9 +74,13 @@ result = app.invoke({"input": "I was billed twice for last month's subscription.print(result["response"]) # -> "Handling billing..."```-`confidence_threshold` reads `answer_confidence`, the calibrated `max(p)` confidence the-calibration figures describe, when the answer carries it, and falls back to the entropy-`confidence` otherwise.+`confidence_threshold` reads `answer_confidence`, the `max(p)` confidence the calibration figures+are computed on, when the answer carries it, and falls back to the entropy `confidence` otherwise.+It is not calibrated as shipped: `laya-multilingual` ships `temperature: [1.0, 1.0, 1.0]` and no+`temperature_by_options` at all, and the english checkpoint's `choice:11+` entry is refused by the+loader's own clamp with a warning that says to treat it as uncalibrated. Fit and validate the+threshold on held-out data at the option counts your workload uses -- see the README's+[Calibration](https://github.com/NandhaKishorM/laya#calibration) section.### Routing with the full conversationdiff --git a/docs/llamaindex.md b/docs/llamaindex.mdindex 1c37c96c..63958a9b 100644--- a/docs/llamaindex.md+++ b/docs/llamaindex.md@@ -117,11 +117,19 @@ Both synchronous `query()` and asynchronous `aquery()` are supported.## 4. Confidence Threshold Gating-Like Laya's LangChain integration, `LayaSingleSelector` and `LayaQueryRouter` read calibrated `answer_confidence` (`max(p)`):+Like Laya's LangChain integration, `LayaSingleSelector` and `LayaQueryRouter` read `answer_confidence` (`max(p)`):- **Automatic Fallback:** Specify `fallback_index` (or `fallback_key`) to seamlessly divert uncertain queries to a safe default engine.- **Strict Guarding:** Set `raise_on_low_confidence=True` on `LayaSingleSelector` to raise `LayaLowConfidenceError` when input is ambiguous, allowing caller escalation.+`answer_confidence` is the quantity temperature scaling fits and the one every calibration figure+in the repository is computed on, which is why the threshold reads it rather than the entropy+`confidence`. It is not calibrated as shipped: `laya-multilingual` ships no fitted temperatures at+all (`temperature: [1.0, 1.0, 1.0]`, empty `temperature_by_options`), and english's `choice:11+` is+refused by the loader's own clamp with a warning that says to treat it as uncalibrated. Fit and+validate any threshold on held-out data at the option counts your workload uses -- see the+README's [Calibration](https://github.com/NandhaKishorM/laya#calibration) section.+---## 5. Remote HTTP Deployments