Skip to content

Latest commit

 

History

History
47 lines (30 loc) · 2.54 KB

File metadata and controls

47 lines (30 loc) · 2.54 KB

Catalyst Projections

Projections are versioned public views over admitted or inspectable Catalyst records. They do not create source truth or silently rewrite accepted chunks.

Current Schemas

Projection Version
retrieval catalyst.retrieval.v1
audit catalyst.audit.v1
boundary inspection catalyst.boundaries.v1
candidate evaluation catalyst.candidate_evaluation.v1
parent/child catalyst.parent_child.v1
sentence window catalyst.sentence_window.v1
code catalyst.code.v1
retrieval sanity catalyst.retrieval_sanity.v1
performance benchmark catalyst.performance_benchmark.v1
context recovery benchmark catalyst.context_recovery_benchmark.v1
specialized mode admission catalyst.specialized_mode_admission.v1

Every public projection must include schema_version and projection_kind.

Lossless And Lossy Views

Retrieval records preserve chunk text, source spans, warnings, and relations needed for retrieval. Audit and boundary-inspection records preserve rejected candidates, repairs, violations, warnings, and evidence summaries. Sentence-window and parent/child projections add context views without rewriting accepted chunk identity.

If a projection omits internal detail, it must expose that through warnings, omissions, audit records, or a narrower projection kind.

Retrieval Sanity

Retrieval sanity records compare candidate strategies against held-out fixture evidence. Each strategy result includes hard invariant status, answer-context adequacy, cost, and retrieval_metrics.

The default ranking method is lexical_query_overlap.v1. It reports ranked candidate IDs, relevant candidate IDs, recall_at_1, recall_at_3, and mrr. These fields are diagnostic and cannot admit chunks that fail hard invariants.

Performance Benchmark

Performance benchmark records compare fixture-defined source shapes against chunking strategies. Each strategy result includes source measures, elapsed time, tracemalloc peak memory, chunk count, token total, repair count, strategy identity, and invariant summary.

Timing and memory fields are diagnostic, machine-dependent evidence. They are not admission gates and cannot override hard invariants.

Compatibility

Schema versions are append-only within a version. Removing fields or changing meaning requires a new schema version. Tests should fail any unversioned public projection.

Deprecation

Deprecated projections must remain readable until a replacement is implemented, documented, and covered by contract tests.