ADR-0064: Extraction provider registry and batch executionCopy Anchor Link

• Status: Accepted
• Date: 2026-09-03
• Authors: IBEX Harness team
• Milestone: 3.5.B.2 Cost-tiered extraction execution
• Depends on: ADR-0063, ADR-0020, ADR-0032

ContextCopy Anchor Link

3.5.B.1 shipped the extraction JSON contract in the worker (PR #680 / 078d837). B.2 must call a real LLM, write memories, and record token usage without:

• importing Go packages/provider or services/memory (ADR-0020)
• stacking @traced_task on Celery (spans already wrap IbexTask.__call__)
• inventing pending_review on memories.status
• loading turn text from checkpoints (hashes only, ADR-0032)
• citing ADR-0036 (that number is unpublished)

No GPU/vllm runner exists in CI (ubuntu-latest only).

DecisionCopy Anchor Link

1) Provider ABC in the workerCopy Anchor Link

ExtractionProvider exposes name, supported_models, and extract(system, user) -> ExtractionCall (raw JSON + usage). OpenAI and vLLM share one OpenAI-compatible HTTP client (httpx). vLLM omits Authorization when the key is empty.

Registry is startup-once, keyed by openai | vllm (IBEX_WORKER_EXTRACTION_PROVIDER / EXTRACTION_PROVIDER), fail-closed: openai without OPENAI_API_KEY raises; vllm without IBEX_WORKER_EXTRACTION_BASE_URL raises. No stub fallback.

Default cloud model is gpt-4o-mini (still listed on OpenAI’s public API pricing as of 2026-09-03). Default vLLM model is Qwen2.5-14B-Instruct; operators may substitute Qwen2.5-7B-Instruct when 14B VRAM does not fit. Manual verification only — not run in CI.

2) Confidence policy (a)Copy Anchor Link

Extraction confidence is recorded on the memory row; it does not gate persistence or invent a review status (ADR-0056/0057).

Overall ExtractedMemory.confidence is forwarded as POST /v1/memories confidence. Safety remains Track C (PII, dedup, temporal conflict).

3) Taxonomy duplicationCopy Anchor Link

Worker VALID_LABELS stays local. If migration 000015 / memory VALID_LABELS changes, update services/worker/app/extraction/schema.py in the same PR (issue #681).

4) Batch task and turn sourceCopy Anchor Link

Task ibex.worker.extraction.extract_session_memories (queue=extraction, IbexTask, soft/hard 120/180s). autoretry_for is transport/5xx (ExtractionTransportError) only. org_id is required in kwargs.

Turns arrive in the payload (turn_index, role, content) because checkpoints have no message text. Payload archival is still missing (ADR-0032). Postgres SessionStore is mandatory for extract_session_memories (fail closed without database_url): SELECT/UPDATE ibex_core.sessions with explicit org_id + RLS GUC; skip turn_index <= last_extracted_turn; require status=completed and deleted_at IS NULL. After each turn's HTTP writes succeed, non-atomic UPDATE last_extracted_turn = GREATEST(...) for that turn (not a shared cross-service transaction). Identity-based idempotency keys and crash-window semantics are resolved by ADR-0065 in 3.5.B.3.

Batch JSON: BatchExtractionResult / TurnExtraction. Prompt is EXTRACTION_SYSTEM_PROMPT_V2 plus a batch suffix. Single-turn uses turns length 1.

5) Memory HTTP temporal fieldsCopy Anchor Link

CreateMemoryRequest gained optional valid_from / valid_until (already on CreateMemoryCommand). Mapping is pass-through. org_id is never accepted in the body.

6) ClickHouse llm_tracesCopy Anchor Link

Worker HTTP insert uses the same columns as Go TraceRecord. org_id is required (refuse). Never store prompt/completion. Empty CLICKHOUSE_DSN skips insert and increments ibex_worker_extraction_clickhouse_skip_total. httpx is used instead of clickhouse-connect so LLM mocks and CH mocks share one HTTP stack; stdlib urllib lacks MockTransport and first-class timeout typing.

7) Cost estimate (dated, not ADR-0036)Copy Anchor Link

OpenAI public prices for gpt-4o-mini as of 2026-09-03: $0.15 / 1M input tokens, $0.60 / 1M output tokens. A representative 10-turn session-close call (~1.2k input + 400 output) is about $0.00042 — same order of magnitude as the old “$0.0008/10-turn” sketch, not a guarantee.

ConsequencesCopy Anchor Link

• Extraction HTTP stays in the worker (no packages/ Python LLM library).
• Operators must enqueue turns in the task payload until archival exists.
• vLLM quality is operator-verified; CI only mocks HTTP.
• httpx is a justified worker dependency (timeouts + MockTransport).