AnotiaWang/deep-research-web-ui

AI deep-research agent that turns any question into a cited report: plans searches, reads real sources, verifies evidence. Self-hosted, multi-provider, Docker-ready.

Deep Research Web UI – What It Is

Deep Research Web UI is a web‑application that lets you ask a research question and automatically generates a cited report. It orchestrates large‑language‑model (LLM) calls, web‑search APIs, and page‑scraping to:

plan and run searches,
retrieve source snippets and full‑page text,
verify findings against the original pages,
let you inspect each citation, and
produce a final Markdown/PDF report.

The tool can run entirely in the browser (Client Mode) or on a server/Docker container (Server Mode) where API keys are kept secret.

Key Features (as described in the README)

Feature	What It Means
Safe & Secure – Client Mode	All configuration and API calls stay in the user’s browser; no keys are sent to a remote server.
Realtime feedback	AI responses stream to the UI so you see the answer as it is generated.
Search visualization	The research workflow is shown as a tree, making it easy to follow how queries were refined and which sources were used.
Export	The finished report can be downloaded as Markdown or PDF.
Broad model support	Uses plain text prompts, so it works with many providers (OpenAI‑compatible, SiliconFlow, DeepSeek, Ollama, LiteLLM, etc.).
Docker support	One‑line `docker run` command to host the app yourself.
Server Mode	API keys are supplied via environment variables; end‑users don’t need to enter them.
Evidence inspection & follow‑up	Click a citation to see the source excerpt, then ask the system to re‑search or update that part of the report.
Focused search	Separate query, time‑range and source filters; automatic retry with a revised query if evidence is weak.
Full‑page reading	When a snippet isn’t enough, the app fetches the whole page to extract verbatim evidence, sharing the token budget across the whole run.

How It Works (high‑level flow)

You enter a research question in the UI.
The app creates a search plan and calls a chosen web‑search provider (Tavily, Firecrawl, Google PSE, You.com, Serply, etc.).
Search results are scraped (via the provider’s API or a self‑hosted scraper) to obtain snippets or full‑page text.
An LLM (e.g., GPT‑4o‑mini, DeepSeek, Ollama) processes the evidence, writes a draft report, and cites each finding.
The UI shows a tree view of the search steps and lets you inspect each citation’s source excerpt.
If you need to verify or expand a point, you can “follow up” – the system runs a new focused search and updates only the affected markdown blocks, preserving the original version in history.
When satisfied, you export the report as Markdown or PDF.

Deployment Options

Mode	Typical Use‑case	Setup Highlights
Client Mode	Static hosting (EdgeOne Pages, Netlify, Vercel) where users supply their own API keys.	No server needed; just serve the built static files.
Server Mode (recommended)	Self‑hosted SaaS or internal tool where you keep the API keys secret.	Run the Nuxt SSR server via Docker or a custom build. Environment variables configure AI provider, model, search provider, proxy, etc.

Docker example (Server Mode):

docker run -p 3000:3000 \
  -e NUXT_PUBLIC_SERVER_MODE=true \
  -e NUXT_AI_API_KEY=your-ai-key \
  -e NUXT_WEB_SEARCH_API_KEY=your-search-key \
  -e NUXT_PUBLIC_AI_PROVIDER=openai-compatible \
  -e NUXT_PUBLIC_AI_MODEL=gpt-4o-mini \
  -e NUXT_PUBLIC_WEB_SEARCH_PROVIDER=tavily \
  anotia/deep-research-web:latest

You can also supply a .env file or configure an outbound HTTP/SOCKS proxy (NUXT_PROXY_URL) for geo‑restricted scraping.

Supported Providers (as listed)

AI models: Any OpenAI‑compatible endpoint, plus SiliconFlow, DeepSeek, OpenRouter, Requesty, Ollama, LiteLLM, etc.
Web search / scraping: Tavily (free credits), Firecrawl (cloud or self‑hosted), fastCRW, Google PSE, You.com, Serply.

The README notes that the app uses plain prompts rather than newer structured‑output features, which maximises compatibility across providers.

Who Might Use This?

Researchers or students who need quickly generated, source‑backed reports.
Content creators looking for AI‑assisted literature reviews.
Teams building internal knowledge‑base tools that require citation‑level traceability.
Anyone who wants a self‑hosted alternative to commercial AI research assistants.

License

MIT – permissive open‑source license.

All information above is taken directly from the repository’s README; no additional features are inferred.