trpc.group/trpc-go/trpc-agent-go/examples/graph

Document Processing Workflow Example

This example demonstrates how to build and execute a small document-processing workflow using the trpc-agent-go graph package with GraphAgent and Runner. It showcases:

• Building graphs with StateGraph
• Creating Function nodes、LLM nodes、Tools nodes
• Implementing conditional routing (AddConditionalEdges / AddToolsConditionalEdges)
• Using schema-backed state management (MessagesStateSchema + reducers)
• Creating GraphAgent from a compiled Graph
• Running with Runner and streaming responses

Features

The example implements a document processing pipeline that:

• Preprocesses input documents
• Analyzes document complexity using LLM
• Routes based on complexity (simple vs complex)
• Processes documents differently based on complexity
• Assesses quality of processed content
• Enhances low-quality content using LLM
• Formats final output with statistics

State Model

This example uses the message-oriented schema returned by graph.MessagesStateSchema(), which includes the following keys:

• messages (managed by LLM/tools nodes through reducers)
• user_input
• last_response
• node_responses (per-node textual outputs)

For the example workflow we also track:

• document_length, word_count, complexity_level
• node_execution_history (added via callbacks for formatting stats)

Architecture

The workflow uses a Graph + GraphAgent + Runner architecture:

• Graph: nodes/edges, compiled by StateGraph
• GraphAgent: wraps the graph + executor
• Runner: sessions, event streaming, persistence
• Function/LLM/Tools nodes: core processing primitives
• Conditional routing: route via AddConditionalEdges/AddToolsConditionalEdges

Usage

Run with default examples:

go run .

Run in interactive mode:

go run . -interactive

Use a different model:

go run . -model "gpt-4"

Runtime State Keys

Important keys used by the example:

• messages, user_input, last_response
• document_length, word_count, complexity_level
• node_execution_history (for stats), error_count (optional)

Example Workflow

User Input (via Runner)
      ↓
   Preprocess
      ↓
   Analyze (LLM)
      ↓
 Route by Complexity
     ↙     ↘
Simple     Complex
Process   Summarize (LLM)
     ↘     ↙
   Assess Quality
      ↓
 Route by Quality
     ↙     ↘
  Good     Poor
    ↓     Enhance (LLM)
    ↓        ↓
   Format Output
      ↓
   Final Result

Interactive Mode

In interactive mode, you can:

• Process custom documents by pasting content
• See real-time workflow execution
• View processing statistics
• Type help for available commands
• Type exit to quit

Customization

To customize the workflow:

• Add nodes by implementing NodeFunc
• Modify conditional routing functions
• Extend StateSchema with custom fields/reducers
• Adjust prompts for LLM nodes
• Add tools with function.NewFunctionTool

Requirements

• Go 1.21 or later
• Valid API key and base URL for your model provider (OpenAI-compatible)
• Network connectivity for LLM calls

Tip: if OPENAI_API_KEY is not set, the example prints a hint.