Vector Engine

Role

Searches IFU/510(k) documents stored in OpenAI Vector Stores for information relevant to the user's query. Handles knowledge_base, documentation, and device_definition intents.

Supported Query Types

• •manufacturer_lookup — "Who makes the AXS Catalyst 5?"
• •device_definition — "What is a microcatheter?"
• •contraindications — "What are contraindications for Solitaire?"
• •indications — "What are the indications for Trevo NXT?"
• •guideline_lookup — "What does the IFU say about deployment technique?"
• •trial_summary — "What clinical trials support the Solitaire?"
• •safety_outcomes — "What safety data exists for the AXS Catalyst?"
• •imaging_criteria — "What imaging is required before using Trevo?"
• •patient_eligibility — "Who is eligible for treatment with Solitaire?"
• •source_lookup — "Show me the IFU for Headway 21"

Pipeline

1. •Extract device variant IDs from devices dict (from equipment_extraction)
2. •Build metadata filter: {"type": "containsany", "key": "device_variant_id", "value": [ids]}
3. •Semantic search via OpenAI Vector Stores API (POST /vector_stores/{id}/search)
4. •Score-threshold filtering (MIN_SCORE = 0.4) — drops noisy low-relevance chunks
5. •Sort by score descending, cap at MAX_CHUNKS (10)
6. •Return structured chunks to vector_output_agent

Input Contract

{
  "normalized_query": "Who makes the AXS Catalyst 5?",
  "devices": {"AXS Catalyst 5": {"ids": ["42", "43"], "conical_category": "L2"}},
  "categories": [],
  "classification": {"primary_intent": "knowledge_base"}
}

Output Contract

{
  "status": "complete",
  "engine": "vector_engine",
  "result_type": "vector_search",
  "data": {
    "query": "...",
    "chunks": [{"text": "...", "file_id": "...", "score": 0.92, "attributes": {}}],
    "device_context": {"AXS Catalyst 5": {"ids": ["42", "43"]}},
    "chunk_count": 5,
    "top_score": 0.92
  },
  "classification": {},
  "confidence": 0.9
}

No-Device-ID Fallback

When no device IDs are found (e.g., "What is a microcatheter?"), the engine searches without a metadata filter — pure semantic match against the full document corpus.

Vector Store

• •Provider: OpenAI Vector Stores API (Assistants v2)
• •Store ID: Configured via VECTOR_STORE_ID env var
• •Documents: IFU and 510(k) PDFs, chunked at 1200 tokens with 200-token overlap
• •Metadata: Each chunk tagged with device_variant_id for scoped search