feat(07-02): singleton OpenAIProvider + GenericOpenAIProvider + MAX_AI_CHARS removal

- openai_provider.py: singleton self._client=AsyncOpenAI(...) in __init__ (D-07), _truncate() 60/40 smart truncation (D-13), removed MAX_AI_CHARS constant, changed __init__ signature to include context_chars, removed _client() method - generic_openai_provider.py: new class, subclasses OpenAIProvider, conditional response_format={"type":"json_object"} on supports_json_mode flag (D-01/D-02), imports parse_classification + parse_suggestions from ai.utils (D-02 contract) - ollama_provider.py: added context_chars kwarg with default 8000, passes through - lmstudio_provider.py: added context_chars kwarg with default 8000, passes through - classifier.py: removed MAX_AI_CHARS constant and text[:MAX_AI_CHARS] slices; truncation now handled inside each provider via _truncate() - requirements.txt: bumped anthropic floor to >=0.95.0 (D-03 output_config support)
2026-06-04 18:58:19 +02:00
parent beb5b5e49d
commit 02bcbb9143
6 changed files with 152 additions and 20 deletions
@@ -0,0 +1,103 @@
+"""GenericOpenAIProvider — unified OpenAI-compatible provider for all 8 compat vendors.
+
+Covers: Groq, xAI/Grok, DeepSeek, OpenRouter, Gemini-compat, Mistral-compat,
+        Ollama, LMStudio (D-16/D-17/D-18).
+
+Key design decisions:
+  - Subclasses OpenAIProvider to inherit singleton _client and _truncate (D-07/D-13).
+  - Conditionally passes response_format={"type":"json_object"} based on
+    supports_json_mode flag (D-01) — Gemini preset sets this to False (D-02).
+  - Always parses the raw response with parse_classification / parse_suggestions
+    imported from ai.utils (D-02 last-resort fallback contract — NEVER redefine
+    locally; CLAUDE.md shared module map rule).
+"""
+from __future__ import annotations
+
+from ai.openai_provider import OpenAIProvider
+from ai.utils import parse_classification, parse_suggestions  # D-02 contract
+
+
+class GenericOpenAIProvider(OpenAIProvider):
+    """OpenAI-compatible provider that enforces JSON mode on every call.
+
+    Named presets (Groq, xAI, DeepSeek, OpenRouter, Gemini-compat, Mistral-compat,
+    Ollama, LMStudio) are factory shortcuts in ai/__init__.py that pass the known
+    base_url default for each vendor.
+
+    supports_json_mode=False routes the provider through the parse_classification()
+    fallback path without sending response_format — used for Gemini (D-02).
+    """
+
+    supports_json_mode: bool = True  # class-level default; overridden per instance
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str,
+        base_url: str | None,
+        context_chars: int = 8000,
+        supports_json_mode: bool = True,
+    ):
+        super().__init__(
+            api_key=api_key,
+            model=model,
+            base_url=base_url,
+            context_chars=context_chars,
+        )
+        # Instance-level flag (may differ from class-level default)
+        self.supports_json_mode = supports_json_mode
+
+    async def classify(
+        self,
+        document_text: str,
+        existing_topics: list[str],
+        system_prompt: str,
+    ):
+        topics_str = ", ".join(existing_topics) if existing_topics else "(none yet)"
+        user_msg = (
+            f"Existing topics: [{topics_str}]\n\n"
+            f"Document text:\n{self._truncate(document_text)}"
+        )
+        create_kwargs = dict(
+            model=self._model,
+            max_tokens=1024,
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_msg},
+            ],
+        )
+        if self.supports_json_mode:
+            # D-01: enforce structured JSON output on all supporting providers
+            create_kwargs["response_format"] = {"type": "json_object"}
+        # else: Gemini preset — omit response_format, fall back to parse_classification()
+        response = await self._client.chat.completions.create(**create_kwargs)
+        raw = response.choices[0].message.content or ""
+        # D-02: parse_classification is the last-resort fallback — always called
+        return parse_classification(raw)
+
+    async def suggest_topics(
+        self,
+        document_text: str,
+        system_prompt: str,
+    ) -> list[str]:
+        user_msg = (
+            "Suggest 3-5 topic names for this document. "
+            "Return ONLY valid JSON: {\"suggested_topics\": [\"topic1\", \"topic2\"]}\n\n"
+            f"Document text:\n{self._truncate(document_text)}"
+        )
+        create_kwargs = dict(
+            model=self._model,
+            max_tokens=256,
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_msg},
+            ],
+        )
+        if self.supports_json_mode:
+            create_kwargs["response_format"] = {"type": "json_object"}
+        response = await self._client.chat.completions.create(**create_kwargs)
+        raw = response.choices[0].message.content or ""
+        # D-02: parse_suggestions is the last-resort fallback — always called
+        return parse_suggestions(raw)
+
+    # health_check is inherited from OpenAIProvider — no override needed