What It Does

You type a question. ResearchPilot builds a research plan, executes dozens of targeted web searches in parallel, scores every source for credibility, decides when to go deeper and when to stop, detects contradictions across findings, and assembles a structured report with inline citations.

No manual knobs. The AI determines scope based on your query — you just pick Quick, Thorough, or Deep.

How It Works

Query → Plan → Search → Extract → Assess → Repeat? → Report

1. Plan — A strategic LLM decomposes your query into 2-5 research aspects with sub-questions and priorities
2. Search — Parallel workers execute targeted queries across Firecrawl + Tavily
3. Extract — Findings are pulled from each source with claim, evidence, confidence, and source attribution
4. Assess — Every source gets a credibility score (domain heuristics + LLM assessment for unknowns)
5. Adapt — Information gain is measured after each round. Rich veins get explored deeper. Diminishing returns trigger early stopping
6. Detect — LLM-based contradiction detection flags conflicting claims across sources
7. Gap-fill — Coverage map identifies missing aspects, triggers up to 2 targeted follow-up rounds
8. Report — Outline → section-by-section generation → intro/conclusion → assembly with ranked sources

Architecture

flowchart LR
    Form["📝 Research Form\nQuick / Thorough / Deep"] --> Planner["🧠 Planner\nstrategic LLM"]
    Planner -->|research plan| W1["⚙️ Worker 1\naspect"] & W2["⚙️ Worker 2\naspect"]

    W1 & W2 -->|findings| State["📦 Research State\nimmutable"]
    State --> Pipeline["📄 Report Pipeline\noutline → sections → assemble"]
    Pipeline --> Report(["✅ Structured Report"])

    W1 & W2 -. search .-> Search["🔍 Firecrawl + Tavily"]
    W1 & W2 -. score .-> Cred["🏅 Credibility Scoring"]
    W1 & W2 -. gain .-> Gain["📊 Info Gain / Stop"]
    W1 & W2 -. LLM .-> Models["⚡ fast · smart · strategic"]

    Pipeline -. SSE stream .-> Form

Key Design Decisions

• Immutable state — Research state is a pure data structure. Every mutation returns a new object. No side effects, easy to serialize, easy to debug.
• Adaptive depth — Information gain is measured after each search round. High novelty → keep going. Low novelty → stop. No wasted API calls.
• Tiered models — Fast/cheap models handle extraction (~70% of LLM calls). Smart models write the report. Strategic models plan. Cuts cost 40-60%.
• Abort-aware pipeline — Client disconnect kills all in-flight LLM calls and searches via AbortSignal propagation. No orphaned API spend.
• Structured outputs — Every LLM call uses Zod schemas via the Vercel AI SDK's Output.object. Type-safe from prompt to response.
• Collect-then-merge — Parallel workers return results independently. The orchestrator merges them sequentially after all settle — no race conditions.

Getting Started

Prerequisites

• Bun v1.2+
• A Firecrawl API key (search + scraping)
• An AI provider key: OpenRouter, OpenAI, or Groq

Install

git clone https://github.com/ajayyAI/researchpilot.git
cd researchpilot
bun install
cp .env.example .env

Configure

Edit .env with your API keys:

FIRECRAWL_API_KEY=fc-...        # required — web search + scraping
AI_PROVIDER=openrouter           # openai | openrouter | groq
OPENROUTER_API_KEY=sk-or-...    # key for your chosen provider

Run

bun dev

Open http://localhost:3000. Type a research topic. Pick an effort level. Go.

Configuration

Required

Model Tiers (Optional)

Override the default model for each tier to optimize cost vs. quality:

Search Providers (Optional)

Effort Levels

Users don't configure breadth or depth directly. They pick an effort level and the AI handles the rest:

Tech Stack

Project Structure

src/
├── app/
│   ├── api/research/
│   │   ├── route.ts              # SSE streaming endpoint
│   │   └── feedback/route.ts     # Clarification questions
│   └── page.tsx
├── components/
│   ├── home-page.tsx             # Main page — SSE client, state machine
│   ├── research-form.tsx         # Query input + effort selector
│   ├── research-plan.tsx         # Plan display card
│   ├── research-progress.tsx     # Live progress tracker + log
│   └── research-report.tsx       # Markdown report with copy/download
└── lib/research/
    ├── engine.ts                 # Orchestrator — parallel workers, gap analysis
    ├── workers.ts                # Aspect-level research with adaptive depth
    ├── planner.ts                # Research plan generation (strategic LLM)
    ├── report-pipeline.ts        # Outline → sections → assembly (smart LLM)
    ├── credibility.ts            # Source credibility scoring
    ├── gain.ts                   # Information gain assessment
    ├── search.ts                 # Multi-provider search with dedup + retry
    ├── state.ts                  # Immutable research state
    ├── providers.ts              # Tiered model selection
    ├── types.ts                  # Zod schemas + TypeScript types
    ├── prompts.ts                # System prompts
    └── text-utils.ts             # Token counting, text splitting

How It Compares

Roadmap

• Exa AI integration — Neural/semantic search as third provider for highest-quality retrieval
• Plan approval — Review and edit the research plan before execution
• Research replay — Interactive trace of the full research tree
• Citation verification — Post-generation claim-to-source matching
• Perspective-guided research — Multi-perspective queries (STORM-inspired)
• Interactive sessions — Steer ongoing research in real time
• Export formats — PDF and DOCX alongside markdown
• Research memory — Vector store for cross-session knowledge
• Local-first mode — Ollama + SearXNG for fully offline research
• MCP server — Expose research capabilities as tools for any AI assistant

Development

bun dev              # Dev server with Turbopack
bun run type-check   # TypeScript type checking
bun run ci           # Biome lint + format check
bun run check        # Auto-fix lint/format issues
bun run build        # Production build
bun run validate     # type-check + lint + build

Contributing

Contributions welcome. Open an issue first to discuss what you'd like to change.

1. Fork the repo
2. Create a feature branch (git checkout -b feat/your-feature)
3. Run bun run validate before committing
4. Open a pull request

License

MIT