Resolution Flow

When you send a request to POST /v1/resolve, it passes through a pipeline designed to maximize speed and minimize cost.

The Pipeline

graph TD
    A[Request] --> B{Auth & Rate Limit}
    B -->|Pass| C{Cache Check}
    C -->|Hit| D[Return Cached Result]
    C -->|Miss| E[Memory Bank Search]
    E --> F[Compute]
    F --> G[Cache Update]
    G --> I[Return Computed Result]

    D -.-> J[Billed: 1 request]
    I -.-> K[Billed: 1 request]

1. Authentication & Middleware

The request is validated for:

• Valid API Key/JWT.
• Rate limits (based on your subscription).
• Sufficient request balance.
• Request body size limits.

2. Cache Check

The system checks the cache for a matching result.

• With Standard Cache, identical and similar queries return cached results via fuzzy text matching.
• With Advanced Cache, semantic similarity is also checked — queries that mean the same thing but are phrased differently can still hit the cache.
• Cache Hit: Returns the result immediately. Billed: 1 request.
• Cache Miss: Proceeds to compute. Billed: 1 request.

See Advanced Cache for details on cache modes.

3. Memory Bank Search

If your App has assigned Memory Banks, the system:

• Searches your Memory Banks for similar past examples.
• Retrieves the top matching examples to provide context.

4. Compute

The system processes:

• Your query.
• Your defined tools.
• Retrieved Memory Bank examples (context).
• Optional context string you provided.
• Optional history of recent tool calls.

It resolves the intent to a tool call.

5. Cache Update

The result is stored in the cache for future use. The system also learns a canonical form of the query, so future similar queries can be served from cache without compute.

Optional Fields

Context

You can provide a context string (max 1,000 characters) to give the LLM additional situational information. This helps when the same query could resolve to different tools depending on the situation.

{
  "query": "show me the details",
  "toolsets": ["crm-v1"],
  "context": "User is viewing the orders page"
}

History

You can provide a history array (max 5 entries) of recently-run tools. This helps the LLM understand conversational context for multi-turn interactions.

{
  "query": "add the first one to my cart",
  "toolsets": ["shop-v1"],
  "history": [
    { "tool": "search_products", "query": "red sneakers" }
  ]
}

Each entry has tool (string) and query (string).