Skip to main content

Python SDK

Install

Initialize

Add a rich turn (extract + store)

Retrieve context

Retrieval knobs

  • preset: balanced | factual | recent | graphy
  • or weights: { semantic, bm25, graph_prior, recency_decay, confidence }
  • top_k: candidate count (default 16)

Budgeting

  • Presets: light (≈400), standard (≈1200), heavy (≈2400)
  • add() returns pack = { target_tokens, hard_cap_tokens, used_tokens, dropped[] }

Optional Agent helper

The helper assembles context (preset='factual' with a light preference boost) and composes a grounded prompt. Provide a custom llm_backend to use another LLM.