Edge runtimes

@devilsdev/rag-pipeline-utils/edge is the library without Node.js APIs: one ES module that needs only Web APIs (fetch, crypto.getRandomValues, TextEncoder, streams). Use it where Node.js modules don't exist:

• Vercel Edge Functions
• Cloudflare Workers without nodejs_compat
• Deno and Netlify Edge Functions

On Node.js, and on Workers with nodejs_compat (Cloudflare Workers), use the main entry: it has everything.

CI runs /edge in workerd with no Node.js compatibility, which has fewer APIs than Vercel's Edge runtime. It checks the result against the Node.js build: cache keys and Qdrant point ids hash the same in both (test/edge).

Vercel recommends Node.js

Vercel now recommends its Node.js runtime over Edge for functions, and Next.js 16.3 no longer supports runtime = 'edge'. For a Next.js route, use the main entry on the Node.js runtime. /edge is for Vercel Edge Functions outside Next.js, and for the other runtimes above.

Use it

// api/ask.js, a Vercel Edge Function
import {
  AnthropicConnector,
  LocalEmbedder,
  MemoryRetriever,
  createRagPipeline,
} from "@devilsdev/rag-pipeline-utils/edge";

export const config = { runtime: "edge" };

const docs = [
  { id: "vacation", content: "Employees get twenty vacation days per year." },
  { id: "remote", content: "Remote work is allowed two days a week." },
];
const embedder = new LocalEmbedder();
embedder.train(docs);
const retriever = new MemoryRetriever();
await retriever.store(
  await Promise.all(
    docs.map(async (d) => ({ ...d, vector: await embedder.embed(d.content) })),
  ),
);

export default async function handler(request) {
  const query = new URL(request.url).searchParams.get("q");
  const pipeline = createRagPipeline({
    embedder,
    retriever,
    llm: new AnthropicConnector({ apiKey: process.env.ANTHROPIC_API_KEY }),
  });
  const result = await pipeline.run({ query, options: { citations: true } });
  return Response.json({ answer: result.answer, sources: result.results });
}

On Workers without nodejs_compat, use env.ANTHROPIC_API_KEY: there is no process.env.

What's in it

Area	Exports
Pipeline	createRagPipeline with streaming, timeout, signal, cache, citations, evaluation and telemetry; packContext
Models	AnthropicConnector, CohereConnector, OllamaConnector, createAiSdkAdapter (any AI SDK provider, including OpenAI), createModelRouter
Retrieval	MemoryRetriever, QdrantRetriever, LocalEmbedder, HybridRetriever, BM25Search, reciprocalRankFusion, the rerankers, MemoryCache, RedisCache
Grounding	checkAnswer, CitationTracker, detectHallucinations, source mapping
Guardrails	detectInjection, GuardrailsPipeline and its guards
Evaluation	The scoring functions, PipelineEvaluator, runEvaluation, compareConfigurations, reports
Agentic, costs	AgenticPipeline and its parts, CostCalculator, CostTracker, TokenBudget

It has the same types as the main entry. The bundle is about 140 KB compressed.

Left out, because it needs files, a Node.js server or Node-only packages:

• .ragrc.json loading, the document loaders and FileVectorStore
• PgVectorRetriever (it needs pg, which uses TCP)
• OpenAIConnector: use createAiSdkAdapter with @ai-sdk/openai instead
• The plugin registry, plugin hub and sandbox, DevServer, hot reload and the CLI

Telemetry

The edge build can't load @opentelemetry/api itself, so pass the module in:

import * as api from "@opentelemetry/api";

const pipeline = createRagPipeline({
  embedder,
  retriever,
  llm,
  telemetry: { api },
});

The spans are the same as on Node.js (OpenTelemetry). You need a tracer provider that works in your runtime.

Differences from Node.js

• Logging. The library's logger writes JSON lines with console instead of pino.
• Separate classes. /edge is its own bundle. Don't mix its objects with the main entry's in one process: instanceof checks across the two fail.
• Timing. On Workers, clocks advance only during I/O (Timing).