Documentation
¶
Overview ¶
Package retriever defines the Retriever component interface for fetching relevant documents from a document store given a query.
Overview ¶
A Retriever is the read path of a RAG (Retrieval-Augmented Generation) pipeline. Given a query string it returns the most relevant schema.Document values from an underlying store (vector DB, keyword index, etc.).
Concrete implementations (VikingDB, Milvus, Elasticsearch, …) live in eino-ext:
github.com/cloudwego/eino-ext/components/retriever/
Relationship to Indexer ¶
[Indexer] and Retriever are complementary:
- Indexer writes documents (and their vectors) to the store
- Retriever reads them back
When both use an embedding.Embedder, it must be the same model — vector dimensions must match or similarity scores will be meaningless.
Result Ordering ¶
Results are ordered by relevance score (descending). Scores and other backend metadata are available via schema.Document.MetaData.
See https://www.cloudwego.io/docs/eino/core_modules/components/retriever_guide/
Index ¶
- func GetImplSpecificOptions[T any](base *T, opts ...Option) *T
- type CallbackInput
- type CallbackOutput
- type Option
- func WithDSLInfo(dsl map[string]any) Option
- func WithEmbedding(emb embedding.Embedder) Option
- func WithIndex(index string) Option
- func WithScoreThreshold(threshold float64) Option
- func WithSubIndex(subIndex string) Option
- func WithTopK(topK int) Option
- func WrapImplSpecificOptFn[T any](optFn func(*T)) Option
- type Options
- type Retriever
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func GetImplSpecificOptions ¶ added in v0.3.6
GetImplSpecificOptions extracts implementation-specific options from opts, merging onto base. Call alongside GetCommonOptions inside Retrieve.
Types ¶
type CallbackInput ¶
type CallbackInput struct {
// Query is the query for the retriever.
Query string
// TopK is the top k for the retriever, which means the top number of documents to retrieve.
TopK int
// Filter is the filter for the retriever.
Filter string
// ScoreThreshold is the score threshold for the retriever, eg 0.5 means the score of the document must be greater than 0.5.
ScoreThreshold *float64
// Extra is the extra information for the retriever.
Extra map[string]any
}
CallbackInput is the input for the retriever callback.
func ConvCallbackInput ¶
func ConvCallbackInput(src callbacks.CallbackInput) *CallbackInput
ConvCallbackInput converts the callback input to the retriever callback input.
type CallbackOutput ¶
type CallbackOutput struct {
// Docs is the documents for the retriever.
Docs []*schema.Document
// Extra is the extra information for the retriever.
Extra map[string]any
}
CallbackOutput is the output for the retriever callback.
func ConvCallbackOutput ¶
func ConvCallbackOutput(src callbacks.CallbackOutput) *CallbackOutput
ConvCallbackOutput converts the callback output to the retriever callback output.
type Option ¶
type Option struct {
// contains filtered or unexported fields
}
Option is a call-time option for a Retriever.
func WithDSLInfo ¶
WithDSLInfo wraps the dsl info option.
func WithEmbedding ¶
WithEmbedding wraps the embedder option.
func WithScoreThreshold ¶
WithScoreThreshold wraps the score threshold option.
func WithSubIndex ¶
WithSubIndex wraps the sub index option.
func WrapImplSpecificOptFn ¶ added in v0.3.6
WrapImplSpecificOptFn wraps an implementation-specific option function so it can be passed alongside standard options. For use by Retriever implementors.
type Options ¶
type Options struct {
// Index is the index for the retriever, index in different retriever may be different.
Index *string
// SubIndex is the sub index for the retriever, sub index in different retriever may be different.
SubIndex *string
// TopK is the top k for the retriever, which means the top number of documents to retrieve.
TopK *int
// ScoreThreshold is the score threshold for the retriever, eg 0.5 means the score of the document must be greater than 0.5.
ScoreThreshold *float64
// Embedding is the embedder for the retriever, which is used to embed the query for retrieval .
Embedding embedding.Embedder
// DSLInfo carries backend-specific filter/query expressions. The structure and
// semantics are defined by the underlying store implementation.
DSLInfo map[string]any
}
Options is the options for the retriever.
func GetCommonOptions ¶
GetCommonOptions extracts standard Options from opts, merging onto base. Implementors must call this to honour caller-provided options:
func (r *MyRetriever) Retrieve(ctx context.Context, query string, opts ...retriever.Option) ([]*schema.Document, error) {
options := retriever.GetCommonOptions(&retriever.Options{TopK: &r.defaultTopK}, opts...)
// use options.TopK, options.ScoreThreshold, options.Embedding, etc.
}
type Retriever ¶
type Retriever interface {
Retrieve(ctx context.Context, query string, opts ...Option) ([]*schema.Document, error)
}
Retriever fetches the most relevant documents from a store for a given query.
Retrieve accepts a natural-language query string and returns matching schema.Document values ordered by relevance (most relevant first). Relevance scores and backend-specific metadata are available in schema.Document.MetaData.
When [Options.Embedding] is set, the implementation converts the query to a vector before searching. The embedder must be the same model used at index time — see [indexer.Options.Embedding].
[Options.ScoreThreshold] is a filter, not a sort: documents scoring below the threshold are excluded entirely. [Options.TopK] caps the number of results returned.
Retrieve can be used standalone or added to a Graph via AddRetrieverNode:
retriever, _ := redis.NewRetriever(ctx, cfg)
docs, _ := retriever.Retrieve(ctx, "what is eino?", retriever.WithTopK(5))
graph.AddRetrieverNode("retriever", retriever)
Source Files