Strategies

RAG strategy implementations.

CRAG(retriever, response_synthesizer)

Bases: Strategy

Corrective RAG with relevance-based routing.

Pipeline pattern: Retrieve ↓ Decide ├─ relevant → [Rerank?] → Generate ├─ irrelevant → WebRetrieve → [Rerank?] → Generate └─ ambiguous → FilterChunks → [Rerank?] → Generate

Component usage in this pattern: 1. relevance_evaluator is optional. When configured, it is used in the Decide stage to evaluate retrieved chunks and set retrieval_decision: - relevant when all chunks pass. - irrelevant when no chunks pass. - ambiguous when only part of the chunks pass. It also stores the filtered subset in filtered_chunks. When it is not configured, CRAG falls back to the relevant path. 2. web_retriever is used only in the irrelevant branch as fallback retrieval. If absent, the branch degrades to generating from the original retrieved chunks. 3. reranker is optional and applied inside each generate branch before synthesis, so the selected branch chunks are reranked prior to response generation.

Initializes a new CRAG strategy.

Parameters:

Name	Type	Description	Default
retriever	BaseRetriever	Primary retriever over the main corpus.	required
response_synthesizer	ResponseSynthesizer	Component used to generate the final answer.	required

MapReduceRAG(transformer, retriever, response_synthesizer, fuse_fn='rrf', top_k=5)

Bases: Strategy

Map-reduce based RAG strategy.

The pipeline follows the pattern: Expand query → Retrieve and fuse (map_reduce) → [Reranker?] → Generate.

Attributes:

Name	Type	Description
reranker	BaseReranker | None	Optional reranker component.

Initializes a new MapReduceRAG strategy.

Parameters:

Name	Type	Description	Default
transformer	BaseQueryTransformer	The query transformer component.	required
retriever	BaseRetriever	The retriever component for fetching chunks.	required
response_synthesizer	ResponseSynthesizer	The response synthesizer component.	required
fuse_fn	str | Callable[[list[list[Chunk]]], list[Chunk]]	Fusion strategy name or custom callable. Defaults to "rrf".	'rrf'
top_k	int	Default top-k retrieval count. Defaults to 5.	5

Raises:

Type	Description
ValueError	If any required dependency is None, or top_k is less than 1.

arun(question, top_k=None, **kwargs) async

Async entry point.

Parameters:

Name	Type	Description	Default
question	str	The user query.	required
top_k	int | None	Number of top results to retrieve. If None, uses the constructor default. Defaults to None.	None
**kwargs	Any	Additional retrieval parameters.	{}

Returns:

Type	Description
dict[str, Any]	dict[str, Any]: Final state dict with response.

Raises:

Type	Description
ValueError	If top_k is less than 1.

run(question, top_k=None, **kwargs)

Synchronous entry point.

Parameters:

Name	Type	Description	Default
question	str	The user query.	required
top_k	int | None	Number of top results to retrieve. If None, uses the constructor default. Defaults to None.	None
**kwargs	Any	Additional retrieval parameters.	{}

Returns:

Type	Description
dict[str, Any]	dict[str, Any]: Final state dict with response.

Raises:

Type	Description
ValueError	If top_k is less than 1.

NaiveRAG(retriever, response_synthesizer)

Bases: Strategy

Linear RAG strategy with optional slots.

The pipeline follows the pattern: [QueryExpansionPipeline?] → Retrieve → [RelevanceFilter?] → [Reranker?] → Generate

When query_expansion is set, the bare retriever step is replaced by the QueryExpansionPipeline subgraph.

Attributes:

Name	Type	Description
query_expansion	QueryExpansionPipeline | None	Optional query expansion pipeline.
reranker	BaseReranker | None	Optional reranker component.
relevance_filter	BaseRelevanceFilter | None	Optional relevance filter component.

Initializes a new NaiveRAG strategy.

Parameters:

Name	Type	Description	Default
retriever	BaseRetriever	The retriever component for fetching chunks.	required
response_synthesizer	ResponseSynthesizer	The response synthesizer component for generating responses.	required

Raises:

Type	Description
ValueError	If retriever or response_synthesizer is None.

RoutedStrategy()

Bases: CompositeStrategy

Composite strategy that routes queries to one of multiple branches.

RoutedStrategy supports classifier mode and router mode. Classifier mode evaluates registered classifiers in registration order and selects the first matching branch (first-match semantics). Router mode delegates branch-name selection to a single BaseRouter. The selected branch name is written into state and dispatched via a switch step.

Initializes a new RoutedStrategy.

add_branch(name, classifier=None, branch=None)

Adds a routed branch.

Parameters:

Name	Type	Description	Default
name	str	Unique branch name.	required
classifier	Callable[[str], bool | Awaitable[bool]] | None	Predicate used for classifier-mode route selection. Must be omitted in router mode.	None
branch	Strategy | Pipeline | None	Branch strategy or pipeline.	None

Returns:

Name	Type	Description
RoutedStrategy	'RoutedStrategy'	This strategy instance for fluent chaining.

Raises:

Type	Description
ValueError	If any of the following is true: 1. A branch uses the reserved no-match sentinel name. 2. A branch with the same name has already been registered. 3. Classifier configuration does not match the selected routing mode.

set_router(router)

Configures this strategy to use router mode.

Calling this method again while the strategy is already in router mode replaces the router, preserves registered branches, and invalidates the cached pipeline so the next build uses the new router.

Parameters:

Name	Type	Description	Default
router	BaseRouter	Router used to select a registered branch name.	required

Returns:

Name	Type	Description
RoutedStrategy	'RoutedStrategy'	This strategy instance for fluent chaining.

Raises:

Type	Description
ValueError	If classifier branches are already registered.