https://github.com/assafelovic/gpt-researcher
来自https://github.com/datawhalechina/Agent-Learning-Hub列出的开源项目参考
项目从2023-05开始创建,根据提交记录,请claude code分析了一下,其演变路径如下:
# GPT-Researcher: Evolution & Key Change Points
> Analysed from 2,980 commits spanning May 2023 to May 2026.
---
## Phase 1 — Proof of Concept (May–July 2023)
The project began with a single commit on **May 12, 2023**. The initial architecture was minimal:
- A Python script driving a Selenium/Chrome browser to scrape web pages
- DuckDuckGo as the only search backend
- OpenAI (GPT-4) for synthesis
- Three report types: `outline_report`, `resource_report`, `detailed_report`
- A FastAPI server + plain HTML/JS frontend added within the first two weeks
**Key milestone:** WebSocket real-time streaming of agent progress (July 7), which gave the project its live "thinking out loud" feel that became its identity.
---
## Phase 2 — Community Bootstrap (July–September 2023)
Growth exploded with community PRs almost immediately after launch. Key additions:
- **Auto-agent selection** (Jul 20): LLM dynamically picks the right agent role instead of hardcoding it — Finance, Security Analyst, Business Analyst, etc.
- **Docker support** (Jul 14–31): Containerized Chrome + app for reproducible deployment
- **LangChain integration** (Aug 18): Moved LLM/config to the LangChain abstraction layer
- JS-rendered page scraping support via Selenium (Aug 24)
The project gained ~150 GitHub stars in this phase.
---
## Phase 3 — Multi-Agent Architecture (Oct 2023–Sep 2024)
This is where the project's research depth fundamentally changed.
**~Oct 2023:** Introduction of a `multi_agents` module using **LangGraph** — a hierarchical system where a Master agent orchestrates Writer, Reviewer, Reviser, Editor, and Publisher sub-agents. This enabled long-form, structured research reports.
**~Sep 2024 (PR #875):** The `detailed_report` type was rebuilt around multi-agent subtopic decomposition — each subtopic gets its own parallel research pass before being merged into a coherent final report.
**~Oct 2024 (PR #941):** Introduction of a **Strategic LLM** concept — a separate, more powerful planning model distinct from the fast execution model — allowing cost/quality tradeoffs.
Other expansions in this phase:
- Azure OpenAI, AWS Bedrock, Ollama support
- Multiple search retrievers: Google, Bing, SearxNG, Arxiv, PubMed
- Vector store integration for hybrid local+web research (`langchain_vectorstore` report source)
- Configurable embedding providers
---
## Phase 4 — Frontend Renaissance (Oct–Dec 2024)
The plain HTML frontend was replaced with a full **Next.js application** (PR #898, Oct 2024):
- TypeScript throughout
- Image carousel in reports (Oct 2024, PR #925/#942)
- Chat-with-research capability (post-report Q&A)
- Mobile-responsive design
- Source deduplication and favicon display
- Structured logging system with downloadable logs (Dec 2024)
**Language support** was added (Dec 2024, PR #1026) — reports can now be generated in any language via config.
---
## Phase 5 — Deep Research & Agent Ecosystem (Feb–Jun 2025)
**February 22, 2025** is the single biggest inflection point in the project's history.
**Deep Research mode** (PR #1179/#1195) was introduced as a wholly new research paradigm: iterative, recursive search loops guided by a planning LLM that decides when to dig deeper vs. synthesize. This directly competed with OpenAI's "Deep Research" product.
Simultaneously:
- **Nodriver/Zendriver headless scraper** (Feb 2025) — a Chrome-automation library without Selenium, enabling stealthier and faster scraping
- **FireCrawl scraper** (Feb 2025) — cloud-based scraping for JS-heavy sites
- **MCP Server** (Mar 29, 2025) — GPT-Researcher exposed as a Model Context Protocol server, integrable into Claude, Cursor, etc.
- **React NPM package** (Feb–Mar 2025) — embeddable `<GPTResearcher />` component
- **Domain filtering** (Feb 2025) — restrict research to specific domains
- **Reasoning model support** (Feb–Mar 2025) — o1, o3, Gemini thinking mode, with `reasoning_effort` config
---
## Phase 6 — Provider Proliferation & Observability (Jun–Nov 2025)
The project became a true multi-provider platform:
| Provider added | Date |
|---|---|
| GigaChat | Feb 2025 |
| OpenRouter | Mar 2025 |
| AI/ML API | May 2025 |
| vLLM (local) | May 2025 |
| DashScope (Alibaba) | Jun/Jul 2025 |
| Netmind | Jul 2025 |
| Avian | Feb 2026 |
| Forge | Feb 2026 |
| MiniMax | Mar 2026 |
**LangChain dependency was removed** (Nov 7, 2025, PR #1547 `migration/langchain`) — a major architectural cleanup that reduced startup time and dependency bloat.
**Hallucination evaluation framework** added (Jun 2025) with a judge-based LLM grader.
**PWA support** (Sep 2025) — installable as a mobile app.
**New UI redesign** (Apr 2025, Sep 2025) — two separate premium UI overhauls, the second introducing a sidebar, preferences modal, and polished report rendering.
**LangSmith observability** (Jan 2026) — full tracing of LLM calls.
---
## Phase 7 — Modern Ecosystem (2026)
The project shifted focus toward ecosystem integrations and reliability:
- **AG2 partnership** (Mar 2026) — AutoGen2/AG2 as a multi-agent pipeline alternative
- **Agent discovery endpoint** (Apr 2026) — REST API for external tools to enumerate available agents
- **OpenAlex retriever** (Apr 2026) — academic literature search
- **Xquik X/Twitter retriever** (Apr 2026)
- **Smart context compression** (Feb 2026) — fast-path skips compression for small document sets
- **URL deduplication** (Feb 2026) — prevents re-scraping identical sources
- **Anthropic cost tracking** (May 2026) — real usage metadata from API responses
- **Max tokens raised to 200k** (May 2026) — supporting modern long-output models
- **Tokentoll CI integration** (May 2026) — GitHub Action that analyzes LLM cost impact of PRs
---
## Summary: Architectural Arc
```
2023-05 Simple script: 1 search engine + 1 LLM → report
2023-07 Real-time WebSocket streaming
2023-08 LangChain abstraction layer
2023-10 Multi-agent LangGraph pipeline
2024-09 Strategic LLM / planner separation
2024-10 Next.js frontend + images in reports
2024-12 Language support, structured logging
2025-02 Deep Research (iterative recursive loops)
2025-03 MCP server / embeddable React component
2025-06 MCP UI, FireCrawl, nodriver scrapers
2025-07 Hallucination evals, domain filtering
2025-11 LangChain fully removed
2026-02 vLLM, cost tracking, smart compression
2026-04 Agent discovery, academic retrievers
2026-05 200k token cap, MiniMax, Anthropic cost tracking
```
The project evolved from a single-researcher curiosity in May 2023 into a production-grade, multi-provider autonomous research platform with 1,780+ PRs merged across nearly 200 contributors in just over 3 years.