cat-stack 1.1.0__tar.gz → 1.1.1__tar.gz

{cat_stack-1.1.0 → cat_stack-1.1.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cat-stack
-Version: 1.1.0
+Version: 1.1.1
 Summary: Domain-agnostic text, image, PDF, and DOCX classification engine powered by LLMs
 Project-URL: Documentation, https://github.com/chrissoria/cat-stack#readme
 Project-URL: Issues, https://github.com/chrissoria/cat-stack/issues

{cat_stack-1.1.0 → cat_stack-1.1.1}/src/catstack/__about__.py

@@ -1,7 +1,7 @@
 # SPDX-FileCopyrightText: 2025-present Christopher Soria <chrissoria@berkeley.edu>
 #
 # SPDX-License-Identifier: GPL-3.0-or-later
-__version__ = "1.1.0"
+__version__ = "1.1.1"
 __author__ = "Chris Soria"
 __email__ = "chrissoria@berkeley.edu"
 __title__ = "cat-stack"

{cat_stack-1.1.0 → cat_stack-1.1.1}/src/catstack/_category_analysis.py

@@ -19,13 +19,22 @@ __all__ = ["has_other_category", "check_category_verbosity"]
 _MAX_HEURISTIC_WORDS = 4
 
 # Tier 1: Anchored patterns — safe at any category length.
-# These only match when the keyword IS the category label itself.
+# These only match when the keyword IS the category label itself (or its prefix).
 _ANCHORED_PATTERNS = [
-    re.compile(r"^other\s*$", re.IGNORECASE),         # exact "Other"
-    re.compile(r"^other\s*[:(]", re.IGNORECASE),       # "Other: ...", "Other (..."
-    re.compile(r"^n/?a\s*$", re.IGNORECASE),           # exact "N/A", "NA"
-    re.compile(r"^miscellaneous\s*$", re.IGNORECASE),  # exact "Miscellaneous"
-    re.compile(r"^catch[\s-]?all\s*$", re.IGNORECASE), # exact "catch-all"
+    re.compile(r"^other\s*$", re.IGNORECASE),           # exact "Other"
+    re.compile(r"^other\s*[:(]", re.IGNORECASE),         # "Other: ...", "Other (..."
+    re.compile(r"^n/?a\s*$", re.IGNORECASE),             # exact "N/A", "NA"
+    re.compile(r"^miscellaneous\s*$", re.IGNORECASE),    # exact "Miscellaneous"
+    re.compile(r"^catch[\s-]?all\s*$", re.IGNORECASE),  # exact "catch-all"
+    # Sentiment/opinion contexts — "Neutral" serves as the ambivalent catch-all
+    re.compile(r"^neutral\s*$", re.IGNORECASE),          # exact "Neutral"
+    re.compile(r"^neutral\s*[:(]", re.IGNORECASE),       # "Neutral: ...", "Neutral (..."
+    # Other common ambivalent labels
+    re.compile(r"^ambiguous\s*$", re.IGNORECASE),        # exact "Ambiguous"
+    re.compile(r"^ambiguous\s*[:(]", re.IGNORECASE),     # "Ambiguous: ..."
+    re.compile(r"^unclear\s*$", re.IGNORECASE),          # exact "Unclear"
+    re.compile(r"^unclassified\s*$", re.IGNORECASE),     # exact "Unclassified"
+    re.compile(r"^unknown\s*$", re.IGNORECASE),          # exact "Unknown"
 ]
 
 # Tier 2: Phrase patterns — only applied to short categories (≤ _MAX_HEURISTIC_WORDS).

{cat_stack-1.1.0 → cat_stack-1.1.1}/src/catstack/_providers.py

@@ -772,12 +772,17 @@ def check_ollama_model(model: str, host: str = "localhost", port: int = 11434) -
         True if model is available, False otherwise
     """
     available_models = list_ollama_models(host, port)
-    # Check for exact match or partial match (e.g., "llama3.2" matches "llama3.2:latest")
     model_lower = model.lower()
+    if ":" in model_lower:
+        # User specified an explicit tag (e.g. "qwen2.5:14b") — require exact
+        # match.  An installed "qwen2.5:7b" must NOT satisfy a request for
+        # "qwen2.5:14b"; the previous prefix-match logic let this through,
+        # which caused silent classification failures downstream.
+        return any(m.lower() == model_lower for m in available_models)
+    # User specified just the family (e.g. "llama3.2") — any installed
+    # variant of that family counts (e.g. "llama3.2:latest", "llama3.2:7b").
     return any(
-        model_lower == m.lower() or
-        m.lower().startswith(f"{model_lower}:") or
-        model_lower.startswith(m.lower().split(":")[0])
+        m.lower() == model_lower or m.lower().startswith(f"{model_lower}:")
         for m in available_models
     )