TreeIndex-Go 🌳

Vectorless RAG Framework for Go — No embeddings, no vector DB, just LLM reasoning.

Inspired by PageIndex

English | 中文

---

中文

这是什么？

TreeIndex 是一个纯 Go 实现的 RAG（检索增强生成）框架。它不使用向量数据库，不做文本切片，而是将文档构建为层级树索引，然后用 LLM 推理来定位和检索相关信息。

核心理念

传统 RAG:  文档 → 切片 → 向量嵌入 → 相似度搜索 → 顶部结果
TreeIndex: 文档 → 树索引 → LLM 推理导航 → 精确内容提取

为什么不用向量？

• 向量搜索依赖 embedding 模型的质量
• 小文档/专业领域 embedding 效果差
• 需要额外部署向量数据库（Pinecone/Chroma/Milvus）
• LLM 本身的推理能力足够做精准定位

特性

• 🔶 零外部依赖 — 仅使用 Go 标准库 + net/http 调 LLM API
• 🌳 层级树索引 — 文档结构自然映射为树
• 🧠 推理式检索 — LLM 读树 → 定位 → 提取，不是向量匹配
• 📄 多格式支持 — PDF、Markdown
• 🤖 多 LLM 支持 — OpenAI、Anthropic（任意兼容 API）
• 📦 单二进制 — go build 即可部署
• 💾 索引持久化 — JSON 文件存储，随时加载查询

快速开始

# 编译
go build -o treeindex ./cmd/treeindex

# 快速查看文档树（不需要 LLM）
./treeindex tree document.md

# 构建索引（需要 LLM API key）
export OPENAI_API_KEY=sk-xxx
./treeindex index document.pdf

# 查询
./treeindex query document.treeindex.json "What is the main conclusion?"

# 也支持 Anthropic
export ANTHROPIC_AUTH_TOKEN=sk-ant-xxx
./treeindex index document.md --provider anthropic

命令

命令	说明
treeindex index <file>	构建树索引（含 LLM 摘要）
treeindex tree <file>	快速查看文档树（不需要 LLM）
treeindex query <index> <query>	查询索引文档
treeindex show <index>	显示完整树结构

选项

标志	说明	默认值
--no-summarize	跳过 LLM 摘要生成	false
--merge <len>	合并小节（< len 字符）	200
--top <n>	返回结果数量	5
--provider <p>	LLM 提供商	openai
--format <f>	输出格式 (text/json)	text

作为库使用

package main

import (
    "github.com/seastarbot/treeindex-go/internal/indexer"
    "github.com/seastarbot/treeindex-go/internal/llm"
    "github.com/seastarbot/treeindex-go/internal/retriever"
)

func main() {
    // 创建 LLM provider
    provider, _ := llm.NewProvider(nil)

    // 构建索引
    idx := indexer.New(provider, true)
    tree, _ := idx.IndexFile("document.pdf")
    tree.SaveToFile("index.json")

    // 查询
    ret := retriever.New(provider)
    results, _ := ret.Retrieve(tree, "What is the conclusion?", 5)
    answer, _ := ret.Answer("What is the conclusion?", results)
    fmt.Println(answer)
}

架构

┌──────────┐    ┌──────────┐    ┌──────────┐
│  Parser  │───▶│  Indexer │───▶│  Tree    │
│ PDF/MD   │    │ LLM摘要  │    │ JSON存储 │
└──────────┘    └──────────┘    └──────────┘
                                     │
                                     ▼
                                ┌──────────┐
                                │ Retriever│
                                │ LLM推理  │
                                └──────────┘

许可证

MIT License

---

English

What is this?

TreeIndex is a pure Go implementation of a RAG (Retrieval-Augmented Generation) framework. Instead of vector embeddings and similarity search, it builds a hierarchical tree index from documents and uses LLM reasoning to navigate and retrieve relevant content.

The Idea

Traditional RAG:  Document → Chunks → Embeddings → Vector Search → Top-K
TreeIndex:        Document → Tree Index → LLM Reasoning → Precise Extraction

Why skip vectors?

• Vector search depends heavily on embedding model quality
• Small documents and domain-specific content get poor embeddings
• Requires deploying a vector database (Pinecone/Chroma/Milvus)
• LLMs are already good at understanding document structure and finding information

Features

• 🔶 Zero dependencies — Go stdlib + net/http only
• 🌳 Hierarchical tree index — Document structure naturally maps to a tree
• 🧠 Reasoning retrieval — LLM reads the tree → locates → extracts
• 📄 Multi-format — PDF, Markdown
• 🤖 Multi-LLM — OpenAI, Anthropic (any compatible API)
• 📦 Single binary — Just go build and deploy
• 💾 Persistent indexes — JSON file storage, load and query anytime

Quick Start

# Build
go build -o treeindex ./cmd/treeindex

# Quick tree view (no LLM needed)
./treeindex tree document.md

# Build index (requires LLM API key)
export OPENAI_API_KEY=sk-xxx
./treeindex index document.pdf

# Query
./treeindex query document.treeindex.json "What is the main conclusion?"

# Also supports Anthropic
export ANTHROPIC_AUTH_TOKEN=sk-ant-xxx
./treeindex index document.md --provider anthropic

Commands

Command	Description
treeindex index <file>	Build tree index (with LLM summaries)
treeindex tree <file>	Quick tree view (no LLM needed)
treeindex query <index> <query>	Query an indexed document
treeindex show <index>	Display full tree structure

Options

Flag	Description	Default
--no-summarize	Skip LLM summary generation	false
--merge <len>	Merge small sections (< len chars)	200
--top <n>	Number of results to return	5
--provider <p>	LLM provider (openai/anthropic)	openai
--format <f>	Output format (text/json)	text

As a Library

package main

import (
    "github.com/seastarbot/treeindex-go/internal/indexer"
    "github.com/seastarbot/treeindex-go/internal/llm"
    "github.com/seastarbot/treeindex-go/internal/retriever"
)

func main() {
    // Create LLM provider
    provider, _ := llm.NewProvider(nil)

    // Build index
    idx := indexer.New(provider, true)
    tree, _ := idx.IndexFile("document.pdf")
    tree.SaveToFile("index.json")

    // Query
    ret := retriever.New(provider)
    results, _ := ret.Retrieve(tree, "What is the conclusion?", 5)
    answer, _ := ret.Answer("What is the conclusion?", results)
    fmt.Println(answer)
}

Architecture

┌──────────┐    ┌──────────┐    ┌──────────┐
│  Parser  │───▶│  Indexer │───▶│  Tree    │
│ PDF/MD   │    │ LLM Summ.│    │ JSON     │
└──────────┘    └──────────┘    └──────────┘
                                     │
                                     ▼
                                ┌──────────┐
                                │ Retriever│
                                │ LLM Rsn. │
                                └──────────┘

License

MIT License