Configuration

FlowRAGConfig

The main configuration object passed to createFlowRAG():

interface FlowRAGConfig {
  schema: Schema;
  storage: {
    kv: KVStorage;
    vector: VectorStorage;
    graph: GraphStorage;
  };
  embedder: Embedder;
  extractor: LLMExtractor;
  reranker?: Reranker;
  evaluator?: Evaluator;
  parsers?: DocumentParser[];
  hooks?: FlowRAGHooks;
  observability?: ObservabilityHooks;
  options?: {
    indexing?: IndexingOptions;
    querying?: QueryOptions;
  };
}

IndexingOptions

interface IndexingOptions {
  chunkSize?: number;          // Tokens per chunk (default: 1200)
  chunkOverlap?: number;       // Overlap between chunks (default: 100)
  maxParallelInsert?: number;  // Concurrent documents (default: 2)
  llmMaxAsync?: number;        // Concurrent LLM calls (default: 4)
  embeddingMaxAsync?: number;  // Concurrent embedding calls (default: 16)
  extractionGleanings?: number; // Additional extraction passes (default: 0)
}

IndexOptions

Per-call options for rag.index():

interface IndexOptions {
  force?: boolean;  // Skip hash check, re-process all (default: false)
}

QueryOptions

interface QueryOptions {
  defaultMode?: QueryMode;  // 'hybrid' | 'local' | 'global' | 'naive' (default: 'hybrid')
  maxResults?: number;      // Max results returned (default: 10)
  vectorWeight?: number;    // Weight for vector results (default: 0.7)
  graphWeight?: number;     // Weight for graph results (default: 0.3)
}

Hooks

interface FlowRAGHooks {
  onEntitiesExtracted?: (
    extraction: ExtractionResult,
    context: ExtractionContext,
  ) => Promise<ExtractionResult>;
}

interface ExtractionContext {
  chunkId: string;
  documentId: string;
  content: string;
}

Observability Hooks

interface ObservabilityHooks {
  onLLMCall?: (event: { model: string; duration: number; usage?: TokenUsage }) => void;
  onEmbedding?: (event: { model: string; textsCount: number; duration: number }) => void;
  onSearch?: (event: { query: string; mode: string; resultsCount: number; duration: number }) => void;
}

Hooks are called from both indexing and querying pipelines with timing information. The onLLMCall hook includes token usage when the provider supports it (e.g., OpenAI extractor).

Environment Variables

Variable	Description	Required
GEMINI_API_KEY	Gemini API key	For Gemini provider
OPENAI_API_KEY	OpenAI API key	For OpenAI provider
ANTHROPIC_API_KEY	Anthropic API key	For Anthropic provider
AWS_REGION	AWS region	For Bedrock/S3/OpenSearch
HF_HOME	Custom cache directory for ONNX models	For Local provider (optional)