Techniques and requirements
Choose supported experiments and understand the model or sidecar each needs.
v0.2.1Local CLI · Offline reports
On this page
Inspect the installed catalog
Terminal
ragvi techniques
ragvi techniques --format json| Experiment | Requirements and behavior |
|---|---|
| bm25 | Lexical independent passages; no model |
| semantic / hybrid | Embedding provider; local FastEmbed needs semantic extra. Hybrid uses weighted RRF. |
| rerank | semantic extra and --reranker-model for a local cross-encoder |
| parent-child | Search children, return full originals; no model with BM25 base |
| heading-context | Authentic headings |
| contextual | --prefixes maps original IDs to text hashes and supplied prefixes |
| multi-query | --variants maps question IDs to 1–8 supplied strings; original is included |
| metadata-filter | --filters with typed predicates before ranking |
| mmr | Embedding provider; cosine relevance/diversity selection |
| rewrite / auto-multi-query | --rewrite-model and your transformation provider |
| compress | Deterministic sentence extraction, or an explicit exact-span model selector |
| import-data | --results with external exact-span or registered-unit results |
Compose stages explicitly
BM25, semantic and hybrid are base retrievers. Modifiers default to BM25; --base-retriever changes the base. --representation selects body, heading_context or contextual. Cross-encoding, MMR, rewriting and compression can be configured per suite variant. These are experiment presets and stages, not a claim of independent algorithm count.
Supplied variants and prefixes
variants.json
{"Q1":["pairing reset duration","hold reset button"]}prefixes.json
{"COPY_ORIGINAL_SOURCE_ID":{"text_hash":"COPY_ORIGINAL_TEXT_HASH","prefix":"Device documentation"}}Unknown question IDs fail. Prefixes must cover every original source and match its text hash. Prefixes and supplied query variants should be prepared independently of expected heldout evidence. Prefix text earns no evidence credit.
Was something unclear? Send documentation feedback.