loreguard-cli 0.14.3__tar.gz → 0.14.6__tar.gz

{loreguard_cli-0.14.3 → loreguard_cli-0.14.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loreguard-cli
-Version: 0.14.3
+Version: 0.14.6
 Summary: Local inference client for Loreguard NPCs
 Project-URL: Homepage, https://loreguard.com
 Project-URL: Documentation, https://github.com/beyond-logic-labs/loreguard-cli#readme
@@ -29,7 +29,7 @@ Requires-Dist: rich>=13.0.0
 Requires-Dist: textual>=0.47.0
 Requires-Dist: tf-keras>=2.16.0
 Requires-Dist: torch>=2.0.0
-Requires-Dist: transformers>=4.36.0
+Requires-Dist: transformers>=5.0.0
 Requires-Dist: uvicorn>=0.27.0
 Requires-Dist: websockets>=12.0
 Provides-Extra: build

{loreguard_cli-0.14.3 → loreguard_cli-0.14.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "loreguard-cli"
-version = "0.14.3"
+version = "0.14.6"
 description = "Local inference client for Loreguard NPCs"
 readme = "README.md"
 license = "MIT"
@@ -28,7 +28,7 @@ dependencies = [
     "aiofiles>=24.1.0",
     "rich>=13.0.0",
     "textual>=0.47.0",
-    "transformers>=4.36.0",
+    "transformers>=5.0.0",
     "torch>=2.0.0",
     "fastapi>=0.109.0",
     "uvicorn>=0.27.0",

{loreguard_cli-0.14.3 → loreguard_cli-0.14.6}/src/intent_classifier.py

@@ -40,16 +40,28 @@ class IntentResult:
 # DeBERTa-v3-large is state-of-the-art for zero-shot classification
 DEFAULT_INTENT_MODEL = "MoritzLaurer/DeBERTa-v3-large-zeroshot-v2.0"
 
-# Intent hypothesis templates for zero-shot classification
-# Each intent maps to a hypothesis that DeBERTa will evaluate
-# NOTE: Hypotheses must be specific to avoid misclassification of mixed-intent messages
-# (e.g., "hey, how are you? what's the ISP fee?" should match LIGHT_RETRIEVAL, not WORKING_MEMORY)
-# TODO: Move hypotheses to backend for centralized control (see loreguard-engine issue)
-INTENT_HYPOTHESES = {
-    IntentLabel.NO_RETRIEVAL: "This is a greeting, farewell, or a vague question about availability without asking for any specific information.",
-    IntentLabel.WORKING_MEMORY: "This asks about the person's current life, recent experiences, what they've been up to, or how things are going for them.",
-    IntentLabel.LIGHT_RETRIEVAL: "This asks for a specific fact, number, price, fee, date, location, or procedure that requires looking up information.",
-    IntentLabel.FULL_RETRIEVAL: "This is a complex question requiring analysis of multiple topics or understanding relationships between different pieces of information.",
+# Intent label descriptions for zero-shot classification.
+# Keep these mutually exclusive and concrete to reduce confusion between:
+# - social greetings vs factual requests
+# - working-memory questions vs retrieval questions
+INTENT_LABEL_DESCRIPTIONS = {
+    IntentLabel.NO_RETRIEVAL: (
+        "a greeting, acknowledgement, farewell, or social small talk "
+        "(examples: hi, hello, hey, yo, thanks, okay, bye) without asking "
+        "for specific factual information"
+    ),
+    IntentLabel.WORKING_MEMORY: (
+        "a question about the NPC's own current state, feelings, recent "
+        "experiences, or personal memory that can be answered from working memory"
+    ),
+    IntentLabel.LIGHT_RETRIEVAL: (
+        "a request for one specific factual detail (number, fee, date, name, "
+        "location, status, or single-file fact) that needs light retrieval"
+    ),
+    IntentLabel.FULL_RETRIEVAL: (
+        "a complex request requiring multiple facts, synthesis, planning, "
+        "comparison, or multi-step reasoning across sources"
+    ),
 }
 
 # Promise detection hypothesis for follow-up triggers (ADR-0020)
@@ -156,32 +168,35 @@ class IntentClassifier:
 
         start_time = time.time()
 
-        # Get candidate labels and hypotheses
-        labels = list(INTENT_HYPOTHESES.keys())
-        hypotheses = list(INTENT_HYPOTHESES.values())
+        # Candidate labels are descriptive intent classes.
+        candidate_labels = list(INTENT_LABEL_DESCRIPTIONS.values())
 
-        # Run zero-shot classification
-        # The pipeline will evaluate each hypothesis against the query
+        # Run zero-shot classification.
+        # Using a consistent hypothesis template tends to be more stable than
+        # passing full natural-language hypotheses as labels.
         result = self._classifier(
             query,
-            candidate_labels=hypotheses,
-            hypothesis_template="{}",  # Use hypotheses directly
+            candidate_labels=candidate_labels,
+            hypothesis_template="This user message is {}.",
             multi_label=False,
         )
 
         latency_ms = int((time.time() - start_time) * 1000)
 
-        # Map the winning hypothesis back to intent label
-        winning_hypothesis = result["labels"][0]
+        # Map the winning description back to intent label
+        winning_description = result["labels"][0]
         confidence = result["scores"][0]
 
-        # Find the intent that corresponds to the winning hypothesis
+        # Find the intent that corresponds to the winning description
         intent = IntentLabel.FULL_RETRIEVAL  # Default
-        for label, hypothesis in INTENT_HYPOTHESES.items():
-            if hypothesis == winning_hypothesis:
+        for label, description in INTENT_LABEL_DESCRIPTIONS.items():
+            if description == winning_description:
                 intent = label
                 break
 
+        # Log full score distribution for tuning/debugging.
+        label_score_pairs = list(zip(result["labels"], result["scores"]))
+        logger.debug("Intent score distribution: %s", label_score_pairs)
         logger.info(f"Intent classification: {intent.value} (confidence={confidence:.2f}, latency={latency_ms}ms)")
 
         return IntentResult(

loreguard_cli-0.14.6/tests/test_intent_classifier.py

@@ -0,0 +1,48 @@
+from src.intent_classifier import (
+    INTENT_LABEL_DESCRIPTIONS,
+    IntentClassifier,
+    IntentLabel,
+)
+
+
+class _FakeClassifier:
+    def __init__(self, labels, scores):
+        self.labels = labels
+        self.scores = scores
+        self.calls = []
+
+    def __call__(self, query, **kwargs):
+        self.calls.append((query, kwargs))
+        return {"labels": self.labels, "scores": self.scores}
+
+
+def test_classify_uses_descriptive_labels_and_maps_winner():
+    winner = INTENT_LABEL_DESCRIPTIONS[IntentLabel.NO_RETRIEVAL]
+    second = INTENT_LABEL_DESCRIPTIONS[IntentLabel.LIGHT_RETRIEVAL]
+    fake = _FakeClassifier(labels=[winner, second], scores=[0.88, 0.12])
+
+    classifier = IntentClassifier(model_path="dummy-model")
+    classifier._classifier = fake
+
+    result = classifier.classify("hi")
+
+    assert result.intent == IntentLabel.NO_RETRIEVAL
+    assert result.confidence == 0.88
+    assert result.latency_ms >= 0
+
+    assert len(fake.calls) == 1
+    _, kwargs = fake.calls[0]
+    assert kwargs["candidate_labels"] == list(INTENT_LABEL_DESCRIPTIONS.values())
+    assert kwargs["hypothesis_template"] == "This user message is {}."
+    assert kwargs["multi_label"] is False
+
+
+def test_classify_with_fallback_returns_full_retrieval_on_error():
+    classifier = IntentClassifier(model_path="dummy-model")
+    classifier._classifier = None
+
+    result = classifier.classify_with_fallback("hi")
+
+    assert result.intent == IntentLabel.FULL_RETRIEVAL
+    assert result.confidence == 0.0
+    assert result.latency_ms == 0

{loreguard_cli-0.14.3 → loreguard_cli-0.14.6}/uv.lock

@@ -600,7 +600,7 @@ wheels = [
 
 [[package]]
 name = "loreguard-cli"
-version = "0.14.0"
+version = "0.14.5"
 source = { editable = "." }
 dependencies = [
     { name = "aiofiles" },