loreguard-cli 0.14.1__tar.gz → 0.14.5__tar.gz

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: loreguard-cli
-Version: 0.14.1
+Version: 0.14.5
 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

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/loreguard_entry.py

@@ -3,6 +3,9 @@
 
 Imports src as a proper package so relative imports inside src/__main__.py work.
 """
+import multiprocessing
+multiprocessing.freeze_support()  # Required for PyInstaller on macOS/Windows
+
 from src.__main__ import main
 
 main()

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "loreguard-cli"
-version = "0.14.1"
+version = "0.14.5"
 description = "Local inference client for Loreguard NPCs"
 readme = "README.md"
 license = "MIT"

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/src/intent_classifier.py

@@ -1,4 +1,4 @@
-"""Intent Classification service for adaptive retrieval (ADR-0010).
+ """Intent Classification service for adaptive retrieval (ADR-0010).
 
 This module provides zero-shot intent classification for the NPC dialogue pipeline.
 It uses DeBERTa-v3-large-zeroshot to classify user messages into retrieval strategy categories:
@@ -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-0.14.1 → loreguard_cli-0.14.5}/src/nli.py

@@ -123,6 +123,10 @@ class NLIService:
                     self._model_path,
                     trust_remote_code=True,
                 )
+                # HHEMv2 custom class may lack all_tied_weights_keys (needed by
+                # newer transformers for .to() / .eval()). Patch if missing.
+                if not hasattr(self._model, "_tied_weights_keys"):
+                    self._model._tied_weights_keys = []
                 self._model.to(self._device)
                 self._model.eval()
 

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/src/tui/screens/main.py

@@ -546,11 +546,17 @@ class MainScreen(Screen):
                 intent_classifier = None
 
             # Load dialogue act classifier (filler selection) - run in thread pool
-            self._update_status("Loading dialogue act model...", log=False)
-            self._log("Loading dialogue act classifier...")
             dialogue_act_classifier = None
+            enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "true").lower() == "true"
+            if not enable_dialogue_act:
+                self._log("Dialogue act classifier disabled via LOREGUARD_DIALOGUE_ACT_ENABLED")
+            else:
+                self._update_status("Loading dialogue act model...", log=False)
+                self._log("Loading dialogue act classifier...")
             try:
-                if is_dialogue_act_model_available():
+                if not enable_dialogue_act:
+                    pass  # Skip loading
+                elif is_dialogue_act_model_available():
                     dialogue_act_classifier = DialogueActClassifier()
                     loop = asyncio.get_event_loop()
                     with concurrent.futures.ThreadPoolExecutor() as pool:

{loreguard_cli-0.14.1 → loreguard_cli-0.14.5}/src/tui/screens/running.py

@@ -2,6 +2,7 @@
 
 import asyncio
 import concurrent.futures
+import os
 from typing import TYPE_CHECKING, Optional
 
 from textual.app import ComposeResult
@@ -247,10 +248,17 @@ class RunningScreen(Screen):
 
                 # Load Dialogue Act Classifier
                 dialogue_act_classifier = None
-                self._update_status("dialogue_act", "Dialogue Act", "Loading...", "info")
-                self._log("Loading Dialogue Act classifier...", "info")
+                enable_dialogue_act = os.getenv("LOREGUARD_DIALOGUE_ACT_ENABLED", "true").lower() == "true"
+                if not enable_dialogue_act:
+                    self._update_status("dialogue_act", "Dialogue Act", "Disabled", "info")
+                    self._log("Dialogue act classifier disabled via LOREGUARD_DIALOGUE_ACT_ENABLED", "info")
+                else:
+                    self._update_status("dialogue_act", "Dialogue Act", "Loading...", "info")
+                    self._log("Loading Dialogue Act classifier...", "info")
                 try:
-                    if is_dialogue_act_model_available():
+                    if not enable_dialogue_act:
+                        pass  # Skip loading
+                    elif is_dialogue_act_model_available():
                         dialogue_act_classifier = DialogueActClassifier()
                         loop = asyncio.get_event_loop()
                         with concurrent.futures.ThreadPoolExecutor() as pool:

loreguard_cli-0.14.5/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-0.14.1 → loreguard_cli-0.14.5}/uv.lock

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