Retrieve

Retrieval paths

Start with BM25 or dense flat search and an explicit allowed reference set. Add context packing to control budget and expose selection decisions. Evaluate approximate search or adaptive routing against that baseline using the same source revisions and query set.

Treat indexes as derived projections of shared source units. The dependency planner coordinates updates to those projections and preserves reusable atom representations.

Retrieval shape	Use
Lexical BM25	Exact terminology, identifiers, and code symbols
Dense flat	Exact small-corpus baseline
HNSW / IVF-PQ	Approximate candidates with measured recall loss
MUVERA + MaxSim	Token-level multi-vector candidate generation and exact reranking
Atom ANN + parent aggregation	Retrieve stable atoms, rank their section, assemble context at query time
SparseCL	Same-topic contradiction candidates
Graph propagation	Multi-hop entity-to-passage recall
Context lifecycle	Selectively inject evidence before a model call and plan updates afterward
Evidence context	Source expansion, citation declarations, compaction, and multimodal evidence

Actual ranking snapshot

Query	Relevant document	BM25 rank	Approximate-index observation
SciFact 100	4381486	1	Lexical retrieval succeeds immediately
SciFact 1099	7662206	3	Relevant result follows two distractors
SciFact 1	31715818	>100	Current BM25 misses at evaluation depth

The detailed page exposes complete top-five candidates and overlap with exact search.

• Multi-vector retrieval
• SparseCL contradiction retrieval
• Hypothetical and hierarchical retrieval
• Adaptive retrieval and compression
• Retrieval plans and context envelopes
• Context lifecycle and selective intervention
• Evidence context

See Semantic atoms and retrieval-time chunks for incremental alignment, atom-vector aggregation, and neighbor expansion.