# Domain — extraction-service (MCP + A2A Opportunities)

## Why extraction-service is ideal for MCP

It already provides:

- a single entrypoint (`POST /extract`, `POST /extract/batch`)
- async extraction jobs (`/extract/jobs`)
- model registry
- sidecar health monitoring and circuit breaker
- rate limits + quotas + cache

Agents can use MCP tools to iterate on prompts/tasks safely and repeatably.

## High-value MCP tool proposals

### Core extraction

- `extraction.extract(text, taskId?, modelId?, productId?)`
- `extraction.extractBatch(inputs, modelId?)`

### Async jobs

- `extraction.submitJob(inputs, modelId?, webhookUrl?)`
- `extraction.getJob(jobId)`
- `extraction.listJobs()`

### Observability

- `extraction.sidecarHealth()`
- `extraction.metrics()`
- `extraction.cacheStats()` (backs `GET /extract/cache-stats`)
- `extraction.sidecarMonitoringState()` (backs `GET /extract/monitoring/sidecar`)

### Rate limits / admin utilities

- `extraction.getProductRateLimitStatus(productId?)`
- `extraction.resetProductRateLimit(productId)` (admin)

## Recommended MCP resources

- `extraction.modelRegistry`
- `extraction.taskCatalog`
  - list task IDs used across products (triage, reflection-enrichment, memory-insight, etc.)
- `extraction.promptGuidelines`

## Recommended A2A workflows

### 1) Task design loop

- **TaskDesignerAgent** drafts:
  - task prompt
  - a small set of examples
- **EvalRunnerAgent** runs:
  - `extractBatch` over an eval set
  - compares JSON shape correctness
- **RegressionAgent** checks:
  - no degradation vs previous baseline

### 2) Extraction incident response

- If extraction errors spike:
  - check sidecar health and circuit breaker state
  - reduce per-product rate limits
  - switch modelId (if supported)

## Product integration hotspots

- MindLyst web API routes proxy to extraction-service (`/api/extract` and triage routes).
- Future: other products can standardize on the same tasks and use a shared task registry.