Recreating DeerFlow’s Multi‑Agent Research Pipeline with LangGraphGo in 30 Minutes

Overview

DeerFlow is an open‑source multi‑agent research framework released on 2025‑05‑10. It orchestrates a series of agents (Planner, Researcher, Reporter, optional Podcast) to automate the full research workflow from information gathering to final report generation.

Feature Analysis

Planner : Decomposes a user query into executable steps and creates a research plan.

Researcher : Executes each step using an LLM (currently the model’s internal knowledge base) and appends results to the state.

Reporter : Concatenates all research results, asks the LLM to produce Markdown, converts it to HTML via github.com/gomarkdown/markdown, and stores the final report.

Podcast (optional) : Generates a two‑host dialogue script from the research results and returns JSON plus an HTML snippet with styled speech bubbles.

Architecture Design

The workflow is defined as a LangGraph state graph where each agent corresponds to a node. The high‑level pipeline is:

用户查询 → 规划器 → 研究员 → 报告者 → （可选）播客 → 最终输出

A conditional edge after the reporter checks GeneratePodcast to decide whether to invoke the podcast node.

State Definition

type State struct {
    Request          Request   `json:"request"`
    Plan             []string  `json:"plan"`
    ResearchResults  []string  `json:"research_results"`
    Images           []string  `json:"images"`
    FinalReport      string    `json:"final_report"`
    PodcastScript    string    `json:"podcast_script"`
    GeneratePodcast  bool      `json:"generate_podcast"`
    Step             int       `json:"step"`
}

Graph Structure

// Core graph construction
workflow.AddNode("planner", "...", PlannerNode)
workflow.AddNode("researcher", "...", ResearcherNode)
workflow.AddNode("reporter", "...", ReporterNode)
workflow.AddNode("podcast", "...", PodcastNode)

// Conditional edge after reporter
workflow.AddConditionalEdge("reporter", func(ctx context.Context, state interface{}) string {
    s := state.(*State)
    if s.GeneratePodcast {
        return "podcast"
    }
    return graph.END
})

Agent Implementations

PlannerNode

Function : Accepts state.Request.Query, prompts the LLM to return a JSON‑encoded plan, and sets state.Plan and state.GeneratePodcast.

Robustness : If JSON parsing fails, falls back to line‑by‑line parsing and keyword detection for the podcast flag.

ResearcherNode

Function : Iterates over state.Plan, calls the LLM for each step with a “researcher” prompt, and appends the LLM’s response to state.ResearchResults. The current implementation uses the LLM’s internal knowledge base; state.Images remains nil.

ReporterNode

Function : Concatenates all state.ResearchResults, asks the LLM to produce Markdown, converts it to HTML via github.com/gomarkdown/markdown, and stores the result in state.FinalReport.

Image handling : Detects placeholders like [IMAGE_X:Title] and replaces them with <img src="..." alt="Title"> when image URLs are available.

PodcastNode

Function : Generates a two‑host dialogue script from the research results, returns JSON, and builds an HTML snippet with styled speech bubbles (Host 1 in blue, Host 2 in pink) plus an “Export JSON” button.

Application Entry & Interaction

CLI Mode

# Run a single query
./deerflow "Your research question"

# Example queries
./deerflow "What are the latest advances in quantum computing?"
./deerflow "Explain AI's impact on healthcare"
./deerflow "Current state of renewable energy"

The CLI creates an initial state, invokes the graph, and prints the final HTML report to stdout.

Web Server Mode

Static assets : Embedded via Go’s embed package from the web/ directory.

API : /api/run triggers the graph using Server‑Sent Events (SSE) for real‑time progress; /api/history lists past queries.

Replay system : If a query has been run before, the server streams the stored log at 200 ms intervals to simulate instant results and save tokens.

Logging : logf writes to stdout and, when a channel is present, pushes logs to the SSE stream.

Key Utility Functions

logf

: Dual‑purpose logger for CLI and web modes. getLLM: Wraps LLM initialization using langchaingo/llm (currently OpenAI/Doubao).

Compilation & Deployment

# Navigate to the DeerFlow showcase
cd showcases/deerflow

# Set environment variables
export OPENAI_API_KEY="your-api-key-here"
# Optional for other providers
export OPENAI_API_BASE="https://api.deepseek.com/v1"
export OPENAI_MODEL="deepseek-chat"

# Build the binary
go build -o deerflow .

Running the Application

Web Interface (recommended)

# Start the server
./deerflow

# Open in a browser
http://localhost:8085

Enter a research query.

Watch real‑time progress via SSE.

View the formatted HTML report.

Access research history and replay past queries.

Command‑Line Interface

# Basic usage
./deerflow "Your research question"

Both interfaces produce the same final report; the web UI adds interactive replay and visualisation.

References

[1] langgraphgo – http://lango.rpcx.io

[2] deerflow showcase – https://github.com/smallnest/langgraphgo/tree/main/showcases/deerflow