@kognitivedev/rag

Provider-agnostic retrieval toolkit with chunkers, vector stores, pipeline ingestion, and managed indexing primitives.

For hosted, zero-config knowledge bases, use @kognitivedev/documents pipelines and @kognitivedev/cloud-knowledge-base. This package is the lower-level toolkit for local or custom RAG systems.

Managed cloud knowledge bases use Qdrant-native hybrid search in the backend. PgVectorStore in this package is still supported for lower-level local/custom RAG, but it is not the managed KB default.

Installation

bun add @kognitivedev/rag @kognitivedev/tools zod

Add @kognitivedev/adapter-ai-sdk only if you want the AI SDK embedding bridge.

Quick Start

import { DocumentPipeline, InMemoryVectorStore, RecursiveTextChunker } from "@kognitivedev/rag";
import { AISDKEmbeddingProvider } from "@kognitivedev/adapter-ai-sdk";
import { openai } from "@ai-sdk/openai";

const pipeline = new DocumentPipeline({
  chunker: new RecursiveTextChunker({ chunkSize: 1000 }),
  embedder: new AISDKEmbeddingProvider({
    model: openai.embedding("text-embedding-3-small"),
  }),
  vectorStore: new InMemoryVectorStore(),
});

await pipeline.ingest([{ content: "Your documents here" }]);
const results = await pipeline.search("query", { topK: 5 });
const tool = pipeline.asTool();

Managed Indexing

Use IndexManager when you need a higher-level index abstraction on top of chunk/embed/store primitives.

import { IndexManager, InMemoryVectorStore, RecursiveTextChunker } from "@kognitivedev/rag";

const manager = new IndexManager({
  chunker: new RecursiveTextChunker({ chunkSize: 1000, overlap: 200 }),
  embedder,
  vectorStore: new InMemoryVectorStore(),
});

const index = manager.createIndex({ name: "docs" });
const source = manager.addSource({ indexId: index.id, kind: "file_upload" });

await manager.syncSource({
  indexId: index.id,
  sourceId: source.id,
  filename: "report.md",
  documents: [{ content: "# Report\n\nImportant content" }],
});

Supported retrieval modes:

• chunks
• files_via_metadata
• files_via_content
• auto_routed

Highlights

• 6 chunkers
• 3 vector stores
• Provider-agnostic EmbeddingProvider interface
• pipeline.asTool() returns a Kognitive tool that works with agents and runtime adapters

For PDFs, DOCX, images, OCR-backed ingestion, and dashboard/cloud knowledge bases, use managed pipelines from @kognitivedev/documents.