Case Intelligence

The Case Intelligence system provides AI-driven decision support on the Overview tab of every compliance case. It replaces raw memory data with a synthesized assessment that compares the current case against organizational precedent.

Architecture Overview

Four Information Streams

The endpoint aggregates four distinct information streams into a single response:

Lazy Assessment Generation

The AI assessment uses a lazy generation pattern to avoid blocking the API response:

Cache Invalidation

The intelligence cache is automatically cleared when new data arrives that could change the assessment:

Similar Cases Enrichment

The similar cases pipeline uses a two-stage approach — Letta for semantic similarity, PostgreSQL for enrichment:

Structured Output

The PydanticAI agent produces a validated structured output:

class CaseIntelligenceResult(BaseModel):
    summary: str           # 2-4 sentence assessment
    pattern: Literal["approved", "rejected", "mixed", "insufficient"]
    key_differences: list[str]  # Max 3
    assessment_confidence: float  # 0-1
    similar_count: int
    approved_count: int

Key constraint: The agent is instructed to never recommend a decision — it only surfaces evidence and patterns. This satisfies EU AI Act Article 14 (human oversight for high-risk AI systems).

EU AI Act Compliance (Art. 12)

Every cached assessment includes traceability metadata:

Field	Purpose
model_id	Which LLM produced the assessment (e.g., openai:gpt-5.2)
prompt_id	Which prompt version was used (e.g., case_intelligence_v1)
generated_at	ISO timestamp of generation

This ensures every AI-driven assessment is fully traceable, retrievable, and auditable — a non-negotiable architectural constraint.

Frontend Component

The CaseIntelligenceCard uses progressive disclosure — collapsed by default with a summary line:

🧠 Case Intelligence · 4 similar cases · 3 approved · AI learning (85%)

When expanded, four sections render sequentially:

1. AI Assessment — pattern badge, summary prose, key differences, refresh button
2. Similar Cases — clickable rows with company name, outcome badge, deep links
3. Rules Applied — collapsible rules with auto-learned vs officer-taught distinction
4. Your Corrections — finding confirmations/rejections (hidden when empty)

API Endpoint

GET /api/cases/{workflow_id}/case-intelligence
GET /api/cases/{workflow_id}/case-intelligence?retry=true

See Case Intelligence API for the full response schema.