feat(intelligence): implement evidence extraction with LLM

[rsalus]

Note for implementers: The design outlined here is a guideline, not a strict specification. Feel free to pivot from the suggested approach if you discover a better solution during implementation. Update the issue with your rationale when deviating significantly.

Summary

Implement LLM-based evidence extraction that evaluates clinical data against policy criteria. This is the core reasoning component of the Intelligence Service that determines whether each policy requirement is met, not met, or unclear based on the clinical documentation.

Context

What This Component Does

The evidence extractor sits in the middle of the Intelligence Service reasoning pipeline:

ClinicalBundle + Policy → [Evidence Extractor] → list[EvidenceItem]

Inputs:

• ClinicalBundle (src/models/clinical_bundle.py) - Structured FHIR data containing patient, conditions, observations, procedures, document_texts
• Policy (src/policies/example_policy.py) - Dict with criteria, evidence_patterns, thresholds

Output:

• list[EvidenceItem] - One item per criterion with status (MET/NOT_MET/UNCLEAR), evidence text, source, confidence

Where It Fits

This component is called by the /analyze endpoint (#9) and its output feeds into the form generator (#7).

Tasks

• Create prompt template using design doc §4.4
• Format clinical data for prompt (helper function)
• Format policy criteria for prompt (helper function)
• Call LLM via llm_client.chat_completion()
• Parse JSON response and map to EvidenceItem list
• Implement fallback for LLM unavailability (pattern matching)
• Add unit tests

Business Logic

Criterion Status Determination

Status	When to Use
MET	Clinical data clearly satisfies the requirement
NOT_MET	Clinical data contradicts or fails to meet requirement
UNCLEAR	Insufficient information or ambiguous documentation

Confidence Scoring

• High (0.8-1.0): Clear, explicit documentation
• Medium (0.5-0.79): Implicit evidence requiring interpretation
• Low (0.0-0.49): Weak or ambiguous evidence

Files

File	Action
apps/intelligence/src/reasoning/evidence_extractor.py	Modify
apps/intelligence/src/tests/test_evidence_extractor.py	Modify
apps/intelligence/src/llm_client.py	Reference
apps/intelligence/src/models/pa_form.py	Reference

Dependencies

Issue	Relationship
#5	Soft dependency - determines policy criteria
#8	Provides policy structure
#7	Downstream - uses EvidenceItem output
#9	Integration point - calls this function

Design References

• §4.1 Clinical Reasoning Architecture
• §4.3 Policy Definition
• §4.4 LLM Reasoning Chain - Contains EVIDENCE_EXTRACTION_PROMPT