diff --git a/.planning/phases/07-redo-and-optimize-llm-integration/07-02-SUMMARY.md b/.planning/phases/07-redo-and-optimize-llm-integration/07-02-SUMMARY.md new file mode 100644 index 0000000..c9421cb --- /dev/null +++ b/.planning/phases/07-redo-and-optimize-llm-integration/07-02-SUMMARY.md @@ -0,0 +1,182 @@ +--- +phase: 07-redo-and-optimize-llm-integration +plan: "02" +subsystem: backend/ai-providers +tags: + - ai + - provider-refactor + - singleton-client + - json-mode + - smart-truncation + - registry-pattern + - pydantic + - wave-2 +dependency_graph: + requires: + - "07-01 (system_settings table, HKDF helpers, xfail stubs)" + provides: + - "ProviderConfig Pydantic model + PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE" + - "GenericOpenAIProvider covering all 8 OpenAI-compat vendors" + - "OpenAIProvider singleton _client lifecycle (D-07)" + - "Smart truncation _truncate() 60/40 (D-13)" + - "Registry-based get_provider(config: ProviderConfig) — no if/elif chain" + - "MAX_AI_CHARS removed from openai_provider.py and classifier.py" + - "anthropic SDK floor bumped to >=0.95.0" + - "6 Wave-2 xfail stubs promoted to passing tests" + affects: + - "07-03 (Anthropic output_config — depends on ProviderConfig and registry)" + - "07-04 (Celery retry — depends on classifier.py clean pass-through)" + - "07-05 (Admin AI panel — depends on ProviderConfig for form validation)" +tech_stack: + added: [] + patterns: + - "ProviderConfig(BaseModel) with extra=forbid; context_chars=0 sentinel for PROVIDER_DEFAULTS resolution" + - "PROVIDER_DEFAULTS dict with 10 entries; SUPPORTS_JSON_MODE dict with gemini=False" + - "GenericOpenAIProvider(OpenAIProvider) with conditional response_format kwarg (D-01/D-02)" + - "Singleton self._client = AsyncOpenAI(...) in __init__ (D-07)" + - "_truncate(): first 60% + last 40% of context_chars (D-13)" + - "_REGISTRY dict in ai/__init__.py replaces if/elif chain (O(1) lookup)" + - "D-02 invariant: parse_classification/parse_suggestions always imported from ai.utils" +key_files: + created: + - backend/ai/provider_config.py + - backend/ai/generic_openai_provider.py + modified: + - backend/ai/openai_provider.py + - backend/ai/ollama_provider.py + - backend/ai/lmstudio_provider.py + - backend/ai/__init__.py + - backend/services/classifier.py + - backend/requirements.txt + - backend/tests/test_ai_providers.py +decisions: + - "context_chars=0 as sentinel in ProviderConfig (not 8000) — enables factory to resolve PROVIDER_DEFAULTS via `config.context_chars or defaults['context_chars']`" + - "GenericOpenAIProvider always calls parse_classification() as last-resort regardless of json_mode — D-02 invariant preserved even when json_object is requested" + - "ai/__init__.py imports GenericOpenAIProvider at module load time; no lazy import needed since provider_config.py has no side effects" + - "anthropic floor bumped to >=0.95.0 to unblock Plan 03 output_config usage (A5 from RESEARCH.md)" +metrics: + duration: "~35 minutes" + completed: "2026-06-04" + tasks_completed: 4 + tasks_total: 4 + files_created: 2 + files_modified: 7 +--- + +# Phase 7 Plan 02: Provider Config, GenericOpenAIProvider, and Registry Factory Summary + +ProviderConfig Pydantic model with 10-provider PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE; GenericOpenAIProvider subclassing OpenAIProvider with JSON-mode conditional on supports_json_mode; singleton _client lifecycle; smart truncation; registry-based get_provider(); MAX_AI_CHARS removed from two files; 6 Wave-2 xfail tests promoted. + +## Tasks Completed + +| Task | Description | Commit | Files | +|------|-------------|--------|-------| +| 1 | ProviderConfig + PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE | beb5b5e | backend/ai/provider_config.py | +| 2 | GenericOpenAIProvider + singleton OpenAIProvider + MAX_AI_CHARS removal + ollama/lmstudio context_chars | 02bcbb9 | openai_provider.py, generic_openai_provider.py, ollama_provider.py, lmstudio_provider.py, classifier.py, requirements.txt | +| 3 | Registry-based get_provider(config: ProviderConfig) | 13eef37 | ai/__init__.py, provider_config.py | +| 4 | Promote 6 Wave-2 xfail stubs to passing | 209b156 | tests/test_ai_providers.py | + +## What Was Built + +### Task 1: ProviderConfig + PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE + +Created `backend/ai/provider_config.py` as a pure data module (no provider class imports): + +- `ProviderConfig(BaseModel)` with `extra="forbid"`: provider_id (str, required), api_key (str, default ""), base_url (Optional[str], default None), model (str, default ""), context_chars (int, default 0 — sentinel meaning "use PROVIDER_DEFAULTS") +- `PROVIDER_DEFAULTS: dict[str, dict]` with 10 entries from RESEARCH.md Pattern 2: openai, anthropic, gemini, groq, xai, deepseek, openrouter, mistral, ollama, lmstudio — each with base_url, model, context_chars +- `SUPPORTS_JSON_MODE: dict[str, bool]` with gemini=False (OpenAI compat endpoint does not support `json_object` string form) and True for all others + +### Task 2: GenericOpenAIProvider + Singleton OpenAIProvider + Removals + +**backend/ai/openai_provider.py** refactored: +- `__init__` now accepts `context_chars: int` parameter +- `self._client = AsyncOpenAI(api_key=self._api_key, base_url=self._base_url)` stored as singleton in `__init__` (D-07) +- `def _client(self)` method deleted entirely +- `MAX_AI_CHARS = 8_000` constant deleted +- `def _truncate(self, text: str) -> str` added: returns text unchanged if len <= context_chars, otherwise `text[:head] + "\n[...truncated...]\n" + text[-tail:]` where head = int(context_chars*0.6), tail = context_chars - head (D-13) +- `classify()`, `suggest_topics()`, `health_check()` updated to use `self._client.chat.completions.create(...)` (no parentheses) + +**backend/ai/generic_openai_provider.py** created: +- `class GenericOpenAIProvider(OpenAIProvider)` with `supports_json_mode` instance attribute +- `__init__` accepts `supports_json_mode: bool = True` kwarg, calls `super().__init__(...)` +- `classify()` and `suggest_topics()` conditionally add `response_format={"type":"json_object"}` when `supports_json_mode is True`; omit it for Gemini preset (D-01/D-02) +- Both methods import and call `parse_classification` / `parse_suggestions` from `ai.utils` — D-02 contract enforced via import line +- `health_check()` inherited from OpenAIProvider + +**backend/ai/ollama_provider.py** and **lmstudio_provider.py**: Added `context_chars: int = 8000` parameter, passed through to `super().__init__()`. + +**backend/services/classifier.py**: Removed `MAX_AI_CHARS = 8_000` constant and replaced `text[:MAX_AI_CHARS]` slices with `text` (truncation now inside provider via `_truncate()`). + +**backend/requirements.txt**: `anthropic>=0.26` bumped to `anthropic>=0.95.0` (D-03 output_config support gate for Plan 03). + +### Task 3: Registry-Based Factory + +**backend/ai/__init__.py** rewritten: +- `_REGISTRY: dict[str, type[AIProvider]]` maps 10 provider_ids to classes +- `def get_provider(config: ProviderConfig) -> AIProvider` with typed signature (no raw dict) +- Resolves effective values from PROVIDER_DEFAULTS when config fields are empty/zero +- AnthropicProvider instantiated without base_url (current ctor contract; Plan 03 widens) +- GenericOpenAIProvider gets `supports_json_mode=SUPPORTS_JSON_MODE[config.provider_id]` +- Raises `ValueError(f"Unknown AI provider: {config.provider_id!r}")` for unknown ids + +### Task 4: Six Wave-2 Tests Promoted + +All 6 tests now pass without `@pytest.mark.xfail`: + +1. **test_get_provider_typed**: Registry lookup returns GenericOpenAIProvider for groq/gemini; ValueError for unknown; supports_json_mode correct; _context_chars from PROVIDER_DEFAULTS +2. **test_client_singleton**: `AsyncOpenAI` class called exactly once per provider instance (mocked via patch) +3. **test_generic_openai_json_mode**: `response_format` present in kwargs when supports_json_mode=True; absent when False +4. **test_context_chars_truncation**: Provider with context_chars=100 truncates 500-char input with "[...truncated...]" +5. **test_smart_truncation**: 1000-char limit on 10000-char input → starts with 600 'H's, ends with 400 'T's +6. **test_gemini_fallback_to_parse_classification**: D-02 contract — `parse_classification` called with raw content AND `response_format` absent from API call kwargs + +`test_anthropic_structured_output` remains xfail (Plan 07-03). + +## Verification Results + +- `grep -v '^#' backend/ai/openai_provider.py | grep -c 'MAX_AI_CHARS'` → 0 +- `grep -v '^#' backend/services/classifier.py | grep -c 'MAX_AI_CHARS'` → 0 +- `grep "from ai.utils import parse_classification" backend/ai/generic_openai_provider.py` → match found +- `grep -c "def parse_classification" backend/ai/utils.py` → 1 (file untouched) +- `backend/ai/__init__.py` contains `_REGISTRY` and `def get_provider(config: ProviderConfig)` and all 10 provider keys +- `requirements.txt` contains `anthropic>=0.95.0` +- Full test suite: **1 failed** (pre-existing test_extract_docx ModuleNotFoundError), **363 passed**, **15 xfailed**, **5 skipped** — no new failures; xfailed count down by 6 + +## Deviations from Plan + +### Auto-fixed Issues + +**1. [Rule 1 - Bug] ProviderConfig.context_chars default changed from 8000 to 0** +- **Found during:** Task 3 (test_get_provider_typed assertion failure) +- **Issue:** Task 1 spec said `context_chars: int = 8000` but Task 3 test asserts `result._context_chars == PROVIDER_DEFAULTS["groq"]["context_chars"]` (128000) when `ProviderConfig(provider_id="groq")` is created without specifying context_chars. With default=8000, `8000 or 128000 = 8000` (truthy short-circuit) — test fails. +- **Fix:** Changed `context_chars: int = 0` (sentinel meaning "unset — use PROVIDER_DEFAULTS in factory"). The factory's `config.context_chars or defaults["context_chars"]` then correctly resolves: `0 or 128000 = 128000`. +- **Files modified:** backend/ai/provider_config.py +- **Commit:** 13eef37 + +## Known Stubs + +None — all plan goals implemented; no placeholders. + +## Threat Flags + +No new threat surface introduced. Changes are purely internal provider class refactoring and factory logic — no new API endpoints, no new DB access patterns, no new network paths. T-07-04 (empty api_key) mitigated: factory normalizes `api_key or "not-needed"` before passing to AsyncOpenAI constructor. + +## Self-Check: PASSED + +Files created/modified: + +- [x] backend/ai/provider_config.py — FOUND (ProviderConfig + PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE) +- [x] backend/ai/generic_openai_provider.py — FOUND (GenericOpenAIProvider class present) +- [x] backend/ai/openai_provider.py — FOUND (singleton _client, _truncate, no MAX_AI_CHARS) +- [x] backend/ai/ollama_provider.py — FOUND (context_chars param present) +- [x] backend/ai/lmstudio_provider.py — FOUND (context_chars param present) +- [x] backend/ai/__init__.py — FOUND (_REGISTRY and get_provider(config: ProviderConfig)) +- [x] backend/services/classifier.py — FOUND (MAX_AI_CHARS removed) +- [x] backend/requirements.txt — FOUND (anthropic>=0.95.0) +- [x] backend/tests/test_ai_providers.py — FOUND (6 tests promoted, 1 xfail remaining) + +Commits: +- [x] beb5b5e — feat(07-02): ProviderConfig Pydantic model + PROVIDER_DEFAULTS + SUPPORTS_JSON_MODE +- [x] 02bcbb9 — feat(07-02): singleton OpenAIProvider + GenericOpenAIProvider + MAX_AI_CHARS removal +- [x] 13eef37 — feat(07-02): registry-based get_provider(config: ProviderConfig) — D-06 +- [x] 209b156 — test(07-02): promote 6 Wave-2 xfail tests to passing — D-01/D-02/D-07/D-12/D-13