Files
phase-memory/docs/activation-quality.md

53 lines
1.5 KiB
Markdown

# Activation Quality
`phase-memory` activation can now be planned from deterministic graph
neighborhoods and event paths without requiring embeddings, vector stores, or
LLM ranking.
## Graph Neighborhood Retrieval
`retrieve_graph_neighborhood` expands from seed nodes across graph edges.
Supported controls:
- maximum hops
- edge kind filters
- `in`, `out`, or `both` direction
- phase filters
- memory kind filters
Candidates receive deterministic scores based on explicit priority, graph
distance, phase, lifecycle state, confidence, source-backed status, freshness,
and policy allowance.
## Event Path Activation
`select_event_path` activates ordered event ids from a structured
`MemoryPath`. Active paths are selected by default. Abandoned, merged, or
compacted paths can be included explicitly when a caller wants to inspect
inactive branches.
## Token Estimation
`TokenEstimator` is a protocol. The default implementation is
`WordCountTokenEstimator`, which keeps tests deterministic and dependency-free.
Provider-specific tokenizers can be supplied later without changing retrieval
contracts.
## Quality Report
`activation_quality_report` emits local metrics suitable for later export to
evaluation systems:
- selected expected nodes
- omitted required nodes
- policy-denied required nodes
- token budget utilization
- stale item activation count
- provenance coverage
- source span coverage
- explanation coverage
The fixture `tests/fixtures/activation-quality-report.json` pins a small
expected report for regression tests.