vectorstores

Documentation

Overview

Package vectorstores provides interfaces and implementations for vector databases. Vector stores are used to persist and search document embeddings for RAG applications.

Index

• Variables
• func ToRetriever(vectorStore VectorStore, numDocs int, options ...Option) schema.Retriever
• type ContextNetwork
• type DefinitionRetriever
  • func NewDefinitionRetriever(store VectorStore) (*DefinitionRetriever, error)
  • func (r *DefinitionRetriever) GetDefinition(ctx context.Context, symbolName string) ([]schema.Document, error)
• type DependencyRetriever
  • func NewDependencyRetriever(store VectorStore) (*DependencyRetriever, error)
  • func (r *DependencyRetriever) GetContextNetwork(ctx context.Context, packageName string, imports []string) (*ContextNetwork, error)
• type DocumentWithScore
• type HyDEOption
  • func WithNumGenerations(n int) HyDEOption
• type HyDERetriever
  • func NewHyDERetriever(baseRetriever schema.Retriever, ...) *HyDERetriever
  • func (r *HyDERetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)
• type MultiQueryRetriever
  • func (r MultiQueryRetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)
• type Option
  • func WithCollectionName(name string) Option
  • func WithEmbedder(embedder embeddings.Embedder) Option
  • func WithFilter(key string, value any) Option
  • func WithFilters(filters map[string]any) Option
  • func WithNameSpace(namespace string) Option
  • func WithScoreThreshold(threshold float32) Option
  • func WithSparseQueries(sparse []*schema.SparseVector) Option
  • func WithSparseQuery(sparse *schema.SparseVector) Option
• type Options
  • func ParseOptions(options ...Option) Options
• type RerankingRetriever
  • func (r RerankingRetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)
  • func (r RerankingRetriever) GetRelevantScoredDocuments(ctx context.Context, query string) ([]schema.ScoredDocument, error)
• type Retriever
• type ScoredRetriever
• type VectorStore

Variables

var ErrCollectionNotFound = errors.New("collection not found")

ErrCollectionNotFound is returned when a collection does not exist.

Functions

func ToRetriever

func ToRetriever(vectorStore VectorStore, numDocs int, options ...Option) schema.Retriever

ToRetriever creates a retriever from a vector store.

Types

type ContextNetwork

type ContextNetwork struct {
	// Dependencies are documents that the current code imports/depends on
	Dependencies []schema.Document
	// Dependents are documents that import/depend on the current code (impact analysis)
	Dependents []schema.Document
}

ContextNetwork represents the graph neighborhood of a piece of code

type DefinitionRetriever

type DefinitionRetriever struct {
	// contains filtered or unexported fields
}

func NewDefinitionRetriever

func NewDefinitionRetriever(store VectorStore) (*DefinitionRetriever, error)

func (*DefinitionRetriever) GetDefinition

func (r *DefinitionRetriever) GetDefinition(ctx context.Context, symbolName string) ([]schema.Document, error)

GetDefinition looks up a symbol definition using hybrid search (dense + sparse). Filters by identifier and is_definition metadata for precise results.

type DependencyRetriever

type DependencyRetriever struct {
	// contains filtered or unexported fields
}

DependencyRetriever optimizes RAG by traversing the dependency graph

func NewDependencyRetriever

func NewDependencyRetriever(store VectorStore) (*DependencyRetriever, error)

NewDependencyRetriever creates a new graph-based retriever. It returns an error if store is nil.

func (*DependencyRetriever) GetContextNetwork

func (r *DependencyRetriever) GetContextNetwork(ctx context.Context, packageName string, imports []string) (*ContextNetwork, error)

GetContextNetwork retrieves both upstream dependencies and downstream impact

type DocumentWithScore

type DocumentWithScore struct {
	// Document is the retrieved document.
	Document schema.Document
	// Score is the similarity score (higher is more similar).
	Score float32
}

DocumentWithScore represents a document with its similarity score.

type HyDEOption

type HyDEOption func(*HyDERetriever)

func WithNumGenerations

func WithNumGenerations(n int) HyDEOption

WithNumGenerations sets how many hypothetical documents to generate.

type HyDERetriever

type HyDERetriever struct {
	// BaseRetriever performs the actual similarity search
	BaseRetriever schema.Retriever
	// Generator produces a hypothetical document from a query
	Generator func(ctx context.Context, query string) (string, error)
	// NumGenerations controls how many hypothetical docs to generate (default 1).
	// When > 1, results from all generated docs are deduplicated.
	NumGenerations int
}

HyDERetriever implements the Hypothetical Document Embedding pattern. It asks an LLM to generate a hypothetical answer to the query, then uses that hypothetical answer as the search query against the base retriever. This often finds better matches because the hypothetical answer is closer in embedding space to the actual stored documents.

func NewHyDERetriever

func NewHyDERetriever(baseRetriever schema.Retriever, generator func(ctx context.Context, query string) (string, error), opts ...HyDEOption) *HyDERetriever

func (*HyDERetriever) GetRelevantDocuments

func (r *HyDERetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)

type MultiQueryRetriever

type MultiQueryRetriever struct {
	Store        VectorStore
	LLM          llms.Model
	NumDocuments int
	Count        int
	// Max results to return after deduplication across all query variations.
	// Defaults to NumDocuments when zero.
	MaxResults int
	// Hook to generate sparse vectors for the newly generated queries
	SparseGenFunc func(ctx context.Context, queries []string) ([]*schema.SparseVector, error)
}

func (MultiQueryRetriever) GetRelevantDocuments

func (r MultiQueryRetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)

type Option

type Option func(*Options)

Option configures vector store operations.

func WithCollectionName

func WithCollectionName(name string) Option

WithCollectionName sets the collection name for the operation.

func WithEmbedder

func WithEmbedder(embedder embeddings.Embedder) Option

WithEmbedder sets the embedder for the operation.

func WithFilter

func WithFilter(key string, value any) Option

WithFilter adds a single metadata filter for the search.

func WithFilters

func WithFilters(filters map[string]any) Option

WithFilters sets metadata filters for the search.

func WithNameSpace

func WithNameSpace(namespace string) Option

WithNameSpace sets the namespace for the operation.

func WithScoreThreshold

func WithScoreThreshold(threshold float32) Option

WithScoreThreshold sets the minimum score threshold for results.

func WithSparseQueries

func WithSparseQueries(sparse []*schema.SparseVector) Option

WithSparseQueries sets sparse vectors for batch hybrid search.

func WithSparseQuery

func WithSparseQuery(sparse *schema.SparseVector) Option

WithSparseQuery sets the sparse vector for hybrid search.

type Options

type Options struct {
	// Embedder overrides the default embedder for this operation.
	Embedder embeddings.Embedder
	// NameSpace is an optional namespace for the operation.
	NameSpace string
	// CollectionName specifies the collection to use.
	CollectionName string
	// ScoreThreshold filters results below this score.
	ScoreThreshold float32
	// Filters contains metadata filters for the search.
	Filters map[string]any
	// SparseQuery is the sparse vector for hybrid search.
	SparseQuery *schema.SparseVector
	// SparseQueries are sparse vectors for batch hybrid search.
	SparseQueries []*schema.SparseVector
}

Options contains configuration for vector store operations.

func ParseOptions

func ParseOptions(options ...Option) Options

ParseOptions creates Options from functional options.

type RerankingRetriever

type RerankingRetriever struct {
	Retriever schema.Retriever
	Reranker  schema.Reranker
	TopK      int // Final number of documents to return after reranking

	// Pre-filter candidates before sending to the reranker. Useful for cheap
	// filtering (e.g. BM25 scoring) to reduce the number of documents the
	// expensive LLM-based reranker has to process.
	CandidateFilter func(query string, docs []schema.Document) []schema.Document

	// MinScore filters out documents that have a reranked score below this threshold.
	// If zero, no threshold is applied.
	MinScore float32
}

RerankingRetriever wraps a standard retriever and uses a reranker to refine results.

func (RerankingRetriever) GetRelevantDocuments

func (r RerankingRetriever) GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)

func (RerankingRetriever) GetRelevantScoredDocuments

func (r RerankingRetriever) GetRelevantScoredDocuments(ctx context.Context, query string) ([]schema.ScoredDocument, error)

type Retriever

type Retriever interface {
	GetRelevantDocuments(ctx context.Context, query string) ([]schema.Document, error)
}

Retriever is the interface for fetching relevant documents for a query.

type ScoredRetriever

type ScoredRetriever interface {
	GetRelevantScoredDocuments(ctx context.Context, query string) ([]schema.ScoredDocument, error)
}

ScoredRetriever is a retriever that returns documents with relevance scores.

type VectorStore

type VectorStore interface {
	// AddDocuments adds documents to the vector store and returns their IDs.
	AddDocuments(ctx context.Context, docs []schema.Document, options ...Option) ([]string, error)
	// SimilaritySearch returns documents similar to the query.
	SimilaritySearch(ctx context.Context, query string, numDocuments int, options ...Option) ([]schema.Document, error)
	// SimilaritySearchBatch returns documents similar to multiple queries.
	SimilaritySearchBatch(ctx context.Context, queries []string, numDocuments int, options ...Option) ([][]schema.Document, error)
	// SimilaritySearchWithScores returns documents with similarity scores.
	SimilaritySearchWithScores(ctx context.Context, query string, numDocuments int, options ...Option) ([]DocumentWithScore, error)
	// ListCollections returns all collection names.
	ListCollections(ctx context.Context) ([]string, error)
	// DeleteCollection deletes a collection by name.
	DeleteCollection(ctx context.Context, collectionName string) error
	// DeleteDocumentsByFilter deletes documents matching the filter.
	DeleteDocumentsByFilter(ctx context.Context, filters map[string]any, options ...Option) error
}

VectorStore is the interface for vector database operations. Implementations support document storage, similarity search, and collection management.