Realtime RAG (TypeScript)

Library-first, realtime RAG engine built for production. This is infrastructure, not a chatbot.

You get:

• Realtime ingestion and updates
• pgvector-powered retrieval
• Streaming generation
• Pluggable embeddings and LLMs
• Prisma-friendly store
• Postgres-native LISTEN/NOTIFY
• Node or Bun runtime

Install

npm install realtime-rag @prisma/client

Project layout

realtime-rag/
├─ src/
│  ├─ index.ts
│  ├─ rag/
│  │  ├─ createRAG.ts
│  │  ├─ RAG.ts
│  │  └─ types.ts
│  ├─ ingestion/
│  │  ├─ chunk.ts
│  │  ├─ embed.ts
│  │  └─ ingest.ts
│  ├─ retrieval/
│  │  └─ retrieve.ts
│  ├─ generation/
│  │  └─ stream.ts
│  ├─ stores/
│  │  ├─ pgvector.ts
│  │  └─ memory.ts
│  ├─ realtime/
│  │  └─ postgres.ts
│  ├─ adapters/
│  │  ├─ langchain.ts
│  │  └─ vercel-ai.ts
│  └─ utils/
│     └─ hash.ts
├─ prisma/
│  └─ schema.prisma
├─ tsup.config.ts
├─ package.json
└─ README.md

Core interfaces

export interface Embedder {
  embed(texts: string[]): Promise<number[][]>
}

export interface LLM {
  stream(prompt: string): AsyncIterable<string>
}

export interface VectorStore {
  upsert(vectors: StoredVector[]): Promise<void>
  query(vector: number[], k: number): Promise<StoredVector[]>
  delete(ids: string[]): Promise<void>
}

Quickstart

import { createRAG, memoryStore } from "realtime-rag"

const rag = createRAG({
  embedder,
  llm,
  store: memoryStore()
})

await rag.add({ id: "doc-1", text: "Hello world" })

for await (const token of rag.ask("What is hello?")) {
  process.stdout.write(token)
}

Prisma + pgvector schema

model Embedding {
  id        String   @id
  content   String
  embedding Unsupported("vector(1536)")
  metadata  Json?
  updatedAt DateTime @updatedAt
}

Enable pgvector:

CREATE EXTENSION IF NOT EXISTS vector;
CREATE INDEX embedding_idx ON "Embedding"
USING ivfflat (embedding vector_cosine_ops);

pgvector store

import { PrismaClient } from "@prisma/client"
import { createRAG, pgVectorStore } from "realtime-rag"

const prisma = new PrismaClient()

const rag = createRAG({
  embedder,
  llm,
  store: pgVectorStore(prisma)
})

Realtime sync (Postgres)

import { Client } from "pg"
import { postgresRealtime } from "realtime-rag"

const client = new Client({ connectionString: process.env.DATABASE_URL })
await client.connect()

const realtime = postgresRealtime(client, { channel: "rag_updates" })

Build

npm run build

License

MIT