Expand description
A persistent two-stage late-interaction retrieval engine.
Structs§
- Adaptive
Rerank - Calibration
Entry - Calibration
Key - Calibration
Snapshot - Candidate
Hit - Chunk
- Collections
- Context
Hit - Context
Options - Context
Plan - Field
Stats - Filter
Stats - Hit
- Index
Config - Index
Stats - Logical
Channel - Logical
Plan - MaxSim
Query - A normalized query prepared for repeated MaxSim scoring.
- Multi
Vector Index - Plan
Estimate - Relative work estimate. These units compare plans; they are not latency.
- Planned
Channel - Planner
Stats - Ranking
Signals - Rerank
- Rerank
Plan - Retrieval
Document - Retrieval
Objective - Retrieval
Plan - Retrieval
Response - Retrieval
Trace - Retrieve
Request - Text
Analyzer - Upsert
Document
Enums§
- Calibration
Target - Channel
- Context
Operator - Durability
- Fsync is the default: acknowledge only after segment data, the manifest, and its directory entry are synced. Buffered retains atomic visibility, but does not promise power-loss durability. There is no background flusher.
- Filter
Strategy - Fusion
- Fusion
Operator - Index
Error - Logical
Channel Kind - Logical
Fusion - Physical
Operator - Plan
Reason - Plan
Stage - Predicate
- Quality
Preference - Query
Intent - Representation
- Representation
Kind - Stopwords
- Stop-word removal policy. Only a built-in English list is offered today; custom lists require persisted vocabularies and are future work.
Functions§
- maxsim
- ColBERT’s late-interaction score: sum of per-query-token maxima.
- maxsim_
flat - MaxSim over already-normalized query vectors and a flat document matrix.