Core concept

Implementation context packages

The heart of ImplCache is not “search results.” It is a budgeted implementation package—enough structure for an agent to write correct code against your corpus.

Primary tool

Agents call get_implementation_context with a task description and optional hints such as language, technology, preferred roots, and a max token budget.

{
  "task": "Add a custom command using RegisterCommand and AddMenuItem",
  "language": "cpp",
  "technology": "Example Plugin SDK",
  "preferredRoots": ["example-control-app", "example-plugin-sdk"],
  "maxContextTokens": 2500
}

What a package typically contains

  • requiredApis
  • relevantSymbols
  • sequence
  • examples
  • constraints
  • pitfalls
  • citations
  • coverage and contextFingerprint

The exact fields can evolve before 1.0, but the intent stays the same: actionable implementation context with provenance.

Budgeted on purpose

Token budgets force prioritization. Returning everything related to a topic often hurts more than it helps. ImplCache tries to return the smallest package that is still sufficient for the current task.

Staged depth

  1. get_implementation_context — start here.
  2. find_symbolsearch_knowledge — deepen when needed.
  3. get_document — only when the package is not enough.

This keeps agents from treating “read the corpus” as the default strategy.

Citations and roots

Citations point back into your indexed URI trees. When a query is ambiguous across product families, ImplCache can return needsChoice with available roots so the agent (or user) picks the right corpus instead of guessing.

Related reading: What is ImplCache, ImplCache vs RAG, and the Agent Guide on GitHub.

Ready to try it? Packages ship for Windows, Linux, and Jetson.

Run locally GitHub