Skip to content

API.md

Latest commit

 

History

History
49 lines (38 loc) · 1.81 KB

API.md

File metadata and controls

49 lines (38 loc) · 1.81 KB
Raw

ARPA Micro Series: F1 Mask — API Reference

This document provides a technical specification for the FastAPI Bridge proxy, serving as a zero-latency PII scrubbing middleware.

Request Interface

  • Endpoint: /v1/chat/completions (OpenAI Compatible)
  • Method: POST
  • Payload Schema:
{
    "model": "gpt-4",
    "session_id": "unique-session-id-001",
    "messages": [
        {"role": "user", "content": "Send it to john.doe@acme.com for review."}
    ],
    "temperature": 0.7,
    "stream": false
}

API Behavior

  1. PII Extraction: The Bridge sends the user content to the micro-f1-mask model in Ollama.
  2. Vault Storage: Detected PII (e.g., john.doe@acme.com) is stored in Redis under the session_id.
  3. Scrubbing: The user content is masked (e.g., Send it to [CONTACT_1] for review.).
  4. Cloud Forwarding: The masked prompt is sent to the CLOUD_LLM_URL (OpenAI, Claude, etc.).
  5. Reconstruction: The Cloud LLM response is scanned for tokens (e.g., [CONTACT_1]) and replaced with original values.
  6. Response: The user receives the full reconstructed response.

Session Management

  • Persistence: Mapping is stored in Redis.
  • Isolation: Each session_id has a unique namespace.
  • Consistency: Within a session, the same PII value always maps to the same token.
  • TTL: Default expiration is 7,200 seconds (2 hours).

Configuration (Environment Variables)

Variable Default Description
OLLAMA_URL http://localhost:11434/api/chat Endpoint for PII detection.
F1_MASK_MODEL micro-f1-mask Ollama model name.
CLOUD_LLM_URL https://api.openai.com/v1/chat/completions Target cloud API.
REDIS_HOST localhost Redis host.

Built & Maintained by ARPA Hellenic Logical Systems