Document review & citations, end to end

How a contract becomes reviewed, cited, and trustworthy: the verification cascade, where citations are stored, the playbook and tabular surfaces that ride on it, and the honest edges where verification stops.

The executor workflow

The executor builds a LangGraph StateGraph in api/app/playbooks/executor.py and runs four nodes sequentially against a shared PlaybookExecutionState TypedDict (api/app/playbooks/state.py). The node implementations live in api/app/playbooks/nodes.py.

retrieve → classify → redline → compile → END

Node	What it does	LLM calls
`retrieve`	For each position, runs Postgres FTS (`websearch_to_tsquery`) over the target document's chunks using the position's `detection_keywords`. Top-4 chunks per position (`RETRIEVAL_TOP_K`). Falls back to the document's first chunks when a position has no keywords or FTS returns no hits.	None (pure SQL).
`classify`	One structured-JSON call per position → `{verdict, confidence, matched_fallback_rank, matched_text, cited_chunk_indices, justification}`. Verdict is one of `matches_standard \| matches_fallback \| deviates \| missing`.	N (one per position).
`redline`	For `deviates` verdicts only, one structured-JSON call per deviating position → `{old_text, new_text, justification}` drafted per the position's `redline_strategy`. Non-deviating rows pass through unchanged.	One per `deviates` verdict.
`compile`	Assembles the per-position results into the `playbook_executions.results` JSONB payload, computes the verdict-count summary, and flips the row to `completed` (or `error` if a prior node set `state["error"]`). Sets `completed_at`.	None.