cat-stack 1.6.5__tar.gz → 1.6.7__tar.gz

{cat_stack-1.6.5 → cat_stack-1.6.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cat-stack
-Version: 1.6.5
+Version: 1.6.7
 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-1.6.5 → cat_stack-1.6.7}/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__ = "1.6.5"
+__version__ = "1.6.7"
 __author__ = "Chris Soria"
 __email__ = "chrissoria@berkeley.edu"
 __title__ = "cat-stack"

{cat_stack-1.6.5 → cat_stack-1.6.7}/src/catstack/_providers.py

@@ -125,6 +125,88 @@ _HF_NEEDS_ENABLE_THINKING_OFF = (
 def _hf_model_needs_enable_thinking_off(model: str) -> bool:
     return any(model.startswith(p) for p in _HF_NEEDS_ENABLE_THINKING_OFF)
 
+
+# ---------------------------------------------------------------------------
+# Anthropic deprecated the `temperature` parameter starting with the Opus 4.7 /
+# 4.8 generation: these models return 400 "`temperature` is deprecated for this
+# model." if it is sent. Older models (opus-4-6, sonnet-4-6, sonnet-4-5, and
+# earlier) still accept it. This mirrors the OpenAI reasoning-model handling
+# above — we skip `temperature` up-front for the known-deprecated prefixes in
+# `_build_anthropic_payload`, and `UnifiedLLMClient.complete()` strips it on a
+# runtime 400 as a safety net for future families not yet in this table.
+#
+# Matched by name prefix; extend the tuple when new temperature-free models
+# ship.
+# ---------------------------------------------------------------------------
+_ANTHROPIC_TEMPERATURE_DEPRECATED = (
+    "claude-opus-4-7",
+    "claude-opus-4-8",
+)
+
+
+def _anthropic_supports_temperature(model: str) -> bool:
+    """False for Anthropic models that reject the `temperature` param."""
+    m = (model or "").lower()
+    return not any(m.startswith(p) for p in _ANTHROPIC_TEMPERATURE_DEPRECATED)
+
+
+# ---------------------------------------------------------------------------
+# Ollama reasoning control: per-model-family parameter format for the
+# top-level `think` field on chat / generate requests.
+#
+# Ollama standardized on a single API field name (`think`) but the value
+# type differs per model family — gpt-oss takes an enum, most others take
+# a boolean. See https://docs.ollama.com/capabilities/thinking.
+#
+# Coverage philosophy: list every Ollama reasoning model family we know of
+# AND that uses the `think` field. Reasoning models that gate via other
+# mechanisms (system prompts, chat-template flags) are explicitly noted in
+# the "NOT in registry" comment below and handled elsewhere — adding them
+# here would silently inject a no-op `think` field, which Ollama may
+# accept but won't honor, leading to surprising behavior.
+#
+# Entries are checked longest-prefix-first by `_ollama_think_value()`, so
+# put more-specific prefixes earlier when adding (e.g. `qwen3-coder` before
+# `qwen3` if they differ).
+#
+#   Registry tuple: (model prefix, value-format, low_value, high_value)
+#
+# Models in registry — `think` field works:
+#   gpt-oss          — enum: "low" / "medium" / "high"  (cannot fully disable)
+#   qwen3 / qwen3.*  — bool: True / False               (covers -thinking variants too)
+#   qwq              — bool: True / False               (Qwen QwQ — preceded Qwen3)
+#   deepseek-r1      — bool: True / False               (covers -distill variants)
+#
+# Models NOT in registry — different mechanism, do NOT add here:
+#   magistral        — controlled via system prompt (Mistral Magistral)
+#   exaone-deep      — uses Modelfile-baked reasoning, no API toggle exposed
+#   marco-o1         — uses chat-template wrappers, not `think` field
+#
+# Models with NO reasoning (so `think` should not appear at all):
+#   gemma2/3, llama3.x/4.x, mistral, mistral-nemo, qwen2.5 (non-QwQ),
+#   phi3/4, granite, olmo, codestral, …
+# These are NOT added; the registry's None-return for unmatched prefixes
+# correctly omits the `think` field for them.
+# ---------------------------------------------------------------------------
+_OLLAMA_REASONING_MODELS = (
+    ("gpt-oss",      "enum", "low", "high"),
+    ("qwen3",        "bool", False, True),  # covers qwen3.*, qwen3-*, -thinking-* variants
+    ("qwq",          "bool", False, True),
+    ("deepseek-r1",  "bool", False, True),  # covers -distill-qwen, -distill-llama, etc.
+)
+
+
+def _ollama_think_value(model: str, thinking_budget):
+    """Map cat-stack's thinking_budget to the right Ollama `think` value for
+    this model family. Returns None if the model isn't in the
+    reasoning-capable registry (no `think` field should be set)."""
+    if thinking_budget is None:
+        return None
+    for prefix, fmt, low_val, high_val in _OLLAMA_REASONING_MODELS:
+        if model.startswith(prefix):
+            return low_val if thinking_budget == 0 else high_val
+    return None
+
 __all__ = [
     # Main client
     "UnifiedLLMClient",
@@ -457,6 +539,12 @@ class UnifiedLLMClient:
         elif self.provider in ("huggingface", "huggingface-together"):
             # HuggingFace needs thinking_budget to disable thinking on models that reason by default
             return self._build_openai_payload(messages, json_schema, creativity, force_json, thinking_budget)
+        elif self.provider == "ollama":
+            # Ollama threads thinking_budget to its top-level `think` field for
+            # reasoning-capable models (gpt-oss accepts low/medium/high; others
+            # accept booleans). Without this, gpt-oss family models emit long
+            # <think> blocks by default that bloat per-row generation 3-5x.
+            return self._build_openai_payload(messages, json_schema, creativity, force_json, thinking_budget)
         else:
             # Other OpenAI-compatible providers (xai, mistral, etc.)
             return self._build_openai_payload(messages, json_schema, creativity, force_json)
@@ -532,6 +620,19 @@ class UnifiedLLMClient:
         elif creativity is not None:
             payload["temperature"] = creativity
 
+        # Ollama: per-model-family reasoning control via the top-level
+        # `think` field. gpt-oss expects an enum ("low"/"medium"/"high");
+        # qwen3/deepseek-r1 expect a boolean. Models not in the
+        # `_OLLAMA_REASONING_MODELS` registry don't support reasoning and
+        # get no `think` field (would be a no-op at best, validator-
+        # confusing at worst). Without this, Ollama-served gpt-oss
+        # produces long `<think>` blocks by default that bloat per-row
+        # generation 3-5x.
+        if self.provider == "ollama":
+            think_value = _ollama_think_value(self.model, thinking_budget)
+            if think_value is not None:
+                payload["think"] = think_value
+
         # HuggingFace: disable thinking on model families whose chat
         # template honors `enable_thinking` (Qwen3-family). Other HF-routed
         # models don't need the kwarg, and strict-validator backends
@@ -583,6 +684,14 @@ class UnifiedLLMClient:
         if system_content:
             payload["system"] = system_content
 
+        # Newer Anthropic models (Opus 4.7+) deprecated `temperature` and 400 if
+        # it is sent. Skip it for those known prefixes, and also honor the flag
+        # cached by complete()'s runtime 400 fallback for future families.
+        _temp_ok = (
+            _anthropic_supports_temperature(self.model)
+            and not getattr(self, "_anthropic_temperature_unsupported", False)
+        )
+
         # Extended thinking for Anthropic (minimum 1024 tokens)
         # When thinking is enabled, temperature must be 1 (Anthropic requirement),
         # so we skip setting temperature from creativity in that case
@@ -592,11 +701,12 @@ class UnifiedLLMClient:
                 "type": "enabled",
                 "budget_tokens": budget,
             }
-            payload["temperature"] = 1
+            if _temp_ok:
+                payload["temperature"] = 1
             # When thinking is enabled, max_tokens must be larger than budget_tokens
             if payload["max_tokens"] <= budget:
                 payload["max_tokens"] = budget + 4096
-        elif creativity is not None:
+        elif creativity is not None and _temp_ok:
             payload["temperature"] = creativity
 
         # Use tool calling for structured output (most reliable for Anthropic)
@@ -912,6 +1022,25 @@ class UnifiedLLMClient:
                             payload.pop("reasoning_effort")
                             continue
 
+                    # Anthropic deprecated `temperature` for newer models
+                    # (Opus 4.7+): they 400 with "`temperature` is deprecated
+                    # for this model." Strip it, cache on the client so the
+                    # payload builder skips it for subsequent rows on this
+                    # client, and retry. Safety net for families not yet in
+                    # `_ANTHROPIC_TEMPERATURE_DEPRECATED`.
+                    if (
+                        "temperature" in error_text
+                        and "deprecated" in error_text
+                        and "temperature" in payload
+                    ):
+                        if not getattr(self, '_warned_temperature_deprecated', False):
+                            print(f"\n[CatLLM] Model '{self.model}' deprecated the temperature parameter.")
+                            print(f"  Dropping it and caching for subsequent calls on this client.\n")
+                            self._warned_temperature_deprecated = True
+                        self._anthropic_temperature_unsupported = True
+                        payload.pop("temperature")
+                        continue
+
                     # HuggingFace: try other routers when the current one
                     # rejects the model with a "wrong router" 400.
                     if self._is_hf_wrong_router_400(response.text):

{cat_stack-1.6.5 → cat_stack-1.6.7}/src/catstack/text_functions_ensemble.py

@@ -3043,7 +3043,7 @@ Categorize text responses {cove_categorize}:
                         messages=messages,
                         json_schema=json_schemas[cfg["model"]],
                         creativity=effective_creativity,
-                        thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together") else None,
+                        thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together", "ollama") else None,
                         max_retries=max_retries,
                     )
 
@@ -3100,7 +3100,7 @@ Categorize text responses {cove_categorize}:
                         messages=messages,
                         json_schema=json_schemas[cfg["model"]],
                         creativity=effective_creativity,
-                        thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together") else None,
+                        thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together", "ollama") else None,
                         max_retries=max_retries,
                     )
 
@@ -3184,7 +3184,7 @@ Categorize text responses {cove_categorize}:
                             messages=_retry_messages,
                             json_schema=json_schemas[cfg["model"]],
                             creativity=effective_creativity,
-                            thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together") else None,
+                            thinking_budget=thinking_budget if cfg["provider"] in ("google", "openai", "anthropic", "huggingface", "huggingface-together", "ollama") else None,
                             max_retries=max_retries,
                         )
 
@@ -3638,11 +3638,12 @@ def build_output_dataframes(
     # Populate data
     for idx, result in enumerate(all_results):
         combined_data["input_index"].append(idx)
-        # Truncate input_data for readability
+        # Full input_data (whitespace-collapsed). Truncating here breaks
+        # downstream joins against gold-standard files and silently feeds
+        # truncated text to any pipeline that reuses input_data as input.
         raw = result["response"]
         clean = " ".join(str(raw).split())  # collapse whitespace/newlines
-        preview = clean[:100] + "..." if len(clean) > 100 else clean
-        combined_data["input_data"].append(preview)
+        combined_data["input_data"].append(clean)
         aggregated = result["aggregated"]
 
         # Add PDF metadata if present
@@ -4464,7 +4465,13 @@ def summarize_ensemble(
             # synthesis still has *something* to anchor on (prior behavior).
             original_text_for_synthesis = entry.get("page_text") or page_label
         else:
-            # Truncate input_data for readability; add input_index for joining
+            # Truncate input_data for readability; add input_index for joining.
+            # Truncation is intentional HERE (summarize): inputs can be whole
+            # documents/PDF pages, and full text would bloat the output and the
+            # synthesis context. classify()'s writer (build_output_dataframes)
+            # deliberately does NOT truncate — survey-length inputs there are
+            # reused for downstream joins. See repo TODO for the permanent
+            # per-function input_data design (preview + stable join key).
             clean = " ".join(str(item).split())  # collapse whitespace/newlines
             preview = clean[:100] + "..." if len(clean) > 100 else clean
             row = {