Documentation
¶
Overview ¶
Package rag implements Retrieval-Augmented Generation (RAG) for Koopa.
The rag package provides document indexing and knowledge base integration for LLM applications. It uses Firebase Genkit's PostgreSQL plugin for vector storage and retrieval.
Overview ¶
RAG enhances LLM responses by augmenting prompts with relevant context from a knowledge base. The rag package manages:
- DocStore configuration for vector storage
- Integration with Genkit's PostgreSQL DocStore
Architecture ¶
Genkit PostgreSQL DocStore
|
+-- Vector embedding (via AI Embedder)
+-- Vector storage (PostgreSQL + pgvector)
|
v
Genkit Retriever (ai.Retriever interface)
|
+-- source_type filtering (conversation, file, system)
+-- Semantic search
|
v
LLM (with augmented context)
Key Components ¶
NewDocStoreConfig: Creates configuration for the Genkit PostgreSQL DocStore.
DeleteByIDs: Removes documents by ID for UPSERT emulation.
Source Types ¶
Documents are categorized by source_type:
- SourceTypeConversation: Chat message history
- SourceTypeFile: Indexed file content
- SourceTypeSystem: Built-in system knowledge
Thread Safety ¶
DocStore handles concurrent operations safely.
Index ¶
Constants ¶
const ( // SourceTypeConversation represents chat message history. SourceTypeConversation = "conversation" // SourceTypeFile represents indexed file content. SourceTypeFile = "file" // SourceTypeSystem represents system knowledge (best practices, coding standards). SourceTypeSystem = "system" )
Source type constants for knowledge documents. These define the categories of knowledge stored in the system.
const ( DocumentsTableName = "documents" DocumentsSchemaName = "public" DocumentsIDColumn = "id" DocumentsContentCol = "content" DocumentsEmbeddingCol = "embedding" DocumentsMetadataCol = "metadata" )
Table schema constants for Genkit PostgreSQL plugin. These match the documents table in db/migrations.
const VectorDimension int32 = 768
VectorDimension is the vector dimension used by the pgvector schema. Must match the documents table migration: embedding vector(768). gemini-embedding-001 produces 3072 dimensions by default; we truncate to 768 via OutputDimensionality in EmbedderOptions.
Variables ¶
This section is empty.
Functions ¶
func DeleteByIDs ¶
DeleteByIDs deletes documents by their IDs. Used for UPSERT emulation since Genkit DocStore only supports INSERT. Exported for testing (fuzz tests in rag_test package).
func NewDocStoreConfig ¶
func NewDocStoreConfig(embedder ai.Embedder) *postgresql.Config
NewDocStoreConfig creates a postgresql.Config for the documents table. This factory ensures consistent configuration across production and tests. EmbedderOptions sets OutputDimensionality to match the pgvector schema.
Types ¶
This section is empty.
Source Files