CompeteInsight is an evidence-first AI competitive analysis workspace. Through a multi-agent research pipeline, it turns a rough competitive research question into public sources, structured evidence, red-team-reviewed claims, competitor matrices, Markdown reports, and an AI Analysis Assistant for follow-up questions.

The project is built for the CIS AI-Powered Competitive Analysis Agent Challenge. At this stage, it is optimized more as a public demo and judging workspace than as a full enterprise-grade multi-tenant SaaS product.

• Start a competitive research run from a target product, competitors, research objective, analysis dimensions, and optional seed links.
• The Planning Agent breaks the question down into research dimensions, search queries, source tasks, and quality rules.
• The Search Agent calls public search sources such as Tavily, Exa, Zhihu, and DuckDuckGo fallback to execute batch queries and targeted gap-filling searches.
• The Fetcher retrieves public webpage content and stores auditable SourceDocument records.
• The Evidence Agent extracts structured evidence from real page content, preserving URL, quote, competitor, dimension, confidence, freshness, authority, and relevance.
• The Analyze Agent aggregates evidence into claims, adds red-team risk review, and evaluates research gaps.
• The Report Agent generates the summary, full report, methodology, competitor matrix, recommendations, and export files.
• The AI Analysis Assistant supports follow-up questions based on the current research report, evidence, and claims.
• Each run is stored as local artifacts under data/runs, making the research process traceable and auditable.

The system uses a frontend-backend separated architecture:

• The frontend is a React + Vite single-page workspace responsible for login, research kickoff, Agent Event Log, Artifacts, Report, and AI Assistant interactions.
• The backend is a FastAPI service responsible for authentication, run management, file serving, chat APIs, and agent orchestration.
• ResearchPipeline connects five agent nodes: Planning, Search, Evidence, Analyze, and Report.
• External data comes from Tavily, Exa, Zhihu, and DuckDuckGo fallback. Content extraction is powered by HTTPX, trafilatura, and selectolax.
• The data layer uses a local artifact store, preserving complete research outputs in JSON, JSONL, Markdown, CSV, and related formats.

The pipeline includes a Coverage Gate. When evidence coverage is insufficient, source diversity is too low, claim confidence is weak, or clear counter-evidence risks exist, the system generates gap queries and returns to the Search Agent for targeted gap-filling. Only after the coverage quality gate is satisfied does the pipeline proceed to final report generation.

• Frontend: React 18, TypeScript, Vite, Framer Motion, Lucide icons.
• Backend: FastAPI, Pydantic, Uvicorn, HTTPX.
• LLM: OpenAI-compatible client with support for Ark, DeepSeek, Qwen, and related configurations.
• Search: Tavily, Exa, Zhihu API, DuckDuckGo fallback.
• Content extraction: HTTPX, selectolax, trafilatura.
• Storage: local JSON, JSONL, Markdown, and CSV artifacts.
• Deployment: Nginx, systemd, uv, pnpm.

backend/
  cg/
    agents/          # Agent implementations and runtime helpers
    api/             # FastAPI routers
    llm/             # OpenAI-compatible LLM client
    orchestrator/    # ResearchPipeline orchestration layer
    repositories/    # Local run / evidence repositories
    schemas/         # Pydantic data models
    tools/           # Search and content-fetching tools
frontend/
  src/
    App.tsx
    styles/global.css
skills/              # Skill metadata used by the Agent workspace
scripts/             # Deployment and initialization scripts
diagrams/            # Architecture diagrams used in README and submission docs
data/                # Local run artifacts; production deployment does not overwrite server data

cd backend
uv sync
uv run uvicorn cg.main:app --reload

The default API URL is http://localhost:8000.

cd frontend
pnpm install
pnpm dev

The default frontend URL is http://localhost:5173.

Configure at least one LLM provider in backend/.env.

CG_LLM_PROVIDER=ark
CG_LLM_MODEL=your-model-name
ARK_API_KEY=your-ark-key
ARK_BASE_URL=https://ark.cn-beijing.volces.com/api/v3

TAVILY_API_KEY=your-tavily-key
EXA_API_KEY=your-exa-key
ZHIHU_API_KEY=your-zhihu-key

CG_AUTH_USERNAME=your-demo-user
CG_AUTH_PASSWORD=change-me
CG_AUTH_SECRET=replace-with-a-long-random-secret

Supported LLM API keys include ARK_API_KEY, DEEPSEEK_API_KEY, and QWEN_API_KEY.

The current demo uses a lightweight cookie-session login mechanism:

• POST /api/login
• POST /api/logout
• GET /api/me

cd backend
uv run pytest

cd ../frontend
pnpm build

The hosted demo currently includes a completed sample run for the AI coding assistant landscape:

• Build an Evidence Graph to visualize Source -> Evidence -> Claim -> Recommendation.
• Add report versioning, diff, adoption, and rollback workflows.
• Add research templates for PMs, sales teams, investors, strategy teams, battlecards, and more.