UPAI Open DeepResearch Engine

Standalone open-source DeepResearch engine with:

• adaptive multi-round retrieval
• ranking + domain diversification
• claim clustering + contradiction checks
• citation-grounded synthesis with critique pass
• pluggable adapters for search + LLM providers
• built-in providers for OpenRouter and NVIDIA NIM

Install

npm install @upai/open-deepresearch

Quick Start

import {
  HttpJsonSearchAdapter,
  createOpenDeepResearchEngine,
} from "@upai/open-deepresearch";

const searchAdapter = new HttpJsonSearchAdapter({
  endpoint: process.env.SEARCH_ENDPOINT!,
  apiKey: process.env.SEARCH_API_KEY,
  apiKeyHeader: "Authorization",
  staticPayload: { provider: "free", mode: "web", num: 10 },
});

const engine = createOpenDeepResearchEngine({
  searchAdapter,
  defaultProvider: "openrouter",
  openRouter: {
    apiKey: process.env.OPENROUTER_API_KEY!,
    model: process.env.OPENROUTER_MODEL || "openai/gpt-4o-mini",
  },
  nim: {
    apiKey: process.env.NIM_API_KEY!,
    model: process.env.NIM_MODEL || "meta/llama-3.1-70b-instruct",
  },
  fallbackComplete: async (prompt) => {
    return `Fallback handler received prompt length=${prompt.length}`;
  },
});

const run = await engine.run("Best RAG architecture for support chat in 2026", {
  depth: "deep",
  providerHint: "openrouter",
  onProgress: (p) => console.log(`[${p.stage}] ${p.message}`),
});

console.log(run.finalAnswer);
console.log(run.sourcesForMessage.slice(0, 5));

Required API Keys

At minimum, configure one LLM provider and one search endpoint.

OpenRouter

• OPENROUTER_API_KEY
• OPENROUTER_MODEL (example: openai/gpt-4o-mini)

NVIDIA NIM

• NIM_API_KEY
• NIM_MODEL (example: meta/llama-3.1-70b-instruct)

Search

Use any endpoint that returns JSON with either:

• results: [{ title, snippet, url }]
• OR items: [{ title, snippet, url|link }]

Then configure:

• SEARCH_ENDPOINT
• SEARCH_API_KEY (optional, depending on your backend)

Engine Pipeline

1. Plan: Creates multi-angle query set (facts, recency, benchmarks, counterpoints, docs).
2. Retrieve: Executes search with domain/forum constraints.
3. Rank: Scores by relevance + authority + recency + constraint matching.
4. Diversify: Caps per-domain results to avoid over-concentration.
5. Verify: Builds claim graph and finds likely contradictions.
6. Synthesize: Produces direct answer with citations.
7. Critique: Enforces truth contract and repairs response when possible.

Domain / Forum Constraints

The engine auto-detects constraints from user query language:

• site:reddit.com best vector db
• only from stack overflow
• discussion on hacker news

You can also pass explicit constraints in run(..., { constraints }).

Progress Events

run supports progress hooks for UI and logs:

onProgress: (progress) => {
  // progress.stage: planning | retrieving | ranking | verifying | synthesizing | critiquing
}

Exports

• OpenDeepResearchEngine
• createOpenDeepResearchEngine
• OpenRouterAdapter
• NimAdapter
• HttpJsonSearchAdapter
• FunctionLlmAdapter
• all core types and constraints utilities

Notes

• Node 20+ recommended.
• The library uses fetch and standard Web APIs.
• For strict security, keep provider keys server-side.

License

Apache-2.0