LlamaIndex FrameworkLlamaIndex Framework

模型与开放条件Model & openness

接入方式是「300+ 分散的集成包」，不是统一网关。

• 官方在 integrations 页统称它们 "community integrations"（社区集成）
• 也就是说，多数由社区维护，不是官方自己一个个养

"There are over 300 LlamaIndex integration packages that work seamlessly with core, allowing you to build with your preferred LLM, embedding, and vector store providers."

命名空间规则一眼能记住：

• 带 .core 的（from llama_index.core.xxx）是核心
• 不带的（from llama_index.xxx.yyy）是集成包

新集成要过维护者审核。

README 写明："some integrations may be declined"。

未知：官方未公布集成包的质量分层与维护活跃度分级。 （已查 README 与 integrations 页。）

The integration surface is "300+ separate integration packages", not a unified gateway.

• The official integrations page calls them all "community integrations"
• That is, most are community-maintained, not hand-fed by the vendor one by one

"There are over 300 LlamaIndex integration packages that work seamlessly with core, allowing you to build with your preferred LLM, embedding, and vector store providers."

The namespace rule is easy to remember:

• With .core (from llama_index.core.xxx) = core
• Without it (from llama_index.xxx.yyy) = an integration package

New integrations go through maintainer review.

The README states: "some integrations may be declined".

Unknown: the official docs publish no quality tiering or maintenance-activity grading for the integration packages. (Checked — README and the integrations page.)

运行位置Where it runs

OSS 形态是纯库，有两个安装档位；托管运行在商业平台那一侧。

安装档位有两个（README Getting Started）：

• Starter —— llama-index（PyPI）：核心 + 一批常用集成的合集
• Customized —— llama-index-core：只装核心，再按需从集成页挑包

框架本身是被 import 的库。 接入方式是跟外部应用框架对接， 这也是 README Proposed Solution 第四条：

"integrations with your outer application framework"

官方另有一套托管 / 部署路径。LlamaAgents 的 llamactl CLI 可以这样用：

官方原文："serve locally, and deploy to LlamaCloud or export for self-hosting"

LlamaCloud 另有托管 Index。

• 托管是 LlamaCloud / LlamaAgents 那一层（商业）
• 它不是 OSS 自带的运行形态

The OSS form is a pure library with two install tiers; hosted execution lives on the commercial-platform side.

Two install tiers (README Getting Started):

• Starter — llama-index (PyPI): core plus a bundle of common integrations
• Customized — llama-index-core: core only, then pick packages from the integrations page

The framework itself is a library you import. The way in is to connect with your outer application framework, which is also the fourth item of the README's Proposed Solution:

"integrations with your outer application framework"

The vendor has a separate hosted/deployment path. LlamaAgents' llamactl CLI can be used like this:

Official wording: "serve locally, and deploy to LlamaCloud or export for self-hosting"

LlamaCloud also offers hosted Indexing.

• Hosting is the LlamaCloud / LlamaAgents layer (commercial)
• It is not a run form that ships with the OSS framework

本地文件Local files

它读本地文件是为了建索引，不是给 agent 一个工作区。

官方 SimpleDirectoryReader 文档把这句说得很直白：

"the simplest way to load data from local files into LlamaIndex"

可传的参数：

• input_dir、recursive=True
• exclude、required_exts、num_files_limit
• 远程文件系统走 fsspec（S3、GDrive 等）

索引默认存内存。 两个官方示例动作：

• 落盘：index.storage_context.persist() 写到 ./storage
• 重载：load_index_from_storage

本站记这一轴用的是「索引层视角」。

它不是「agent 工作区」文件系统。 而是「数据源接入 + 索引」层的本地文件读取。

它读盘是为了把文档变成可检索的索引。 这与 Deep Agents 那种「agent 在工作区里改文件」不是一回事。

It reads local files to build indexes, not to give an agent a workspace.

The official SimpleDirectoryReader docs put it plainly:

"the simplest way to load data from local files into LlamaIndex"

Parameters you can pass:

• input_dir, recursive=True
• exclude, required_exts, num_files_limit
• Remote file systems go through fsspec (S3, GDrive, etc.)

Indexes live in memory by default. Two official example actions:

• Persist: index.storage_context.persist() writes to ./storage
• Reload: load_index_from_storage

This site records this axis from the "index layer" perspective.

It is not an "agent workspace" file system. It is local file reading at the "data-source ingestion + indexing" layer.

It reads disk to turn documents into retrievable indexes. That is not the same thing as Deep Agents' "agent editing files in a workspace".

关机后的任务Tasks after shutdown

它没有「关机后继续跑」的托管执行。 但编排（Workflows）确实还在 OSS 里。

有一处容易被 README 误导。 README 把 Workflows 列在 LlamaParse 平台的 Agents 节，官方框架文档却说得很清楚：

"llama-index-core comes with a stable version of Workflows included"

也就是说编排没被移出 OSS。两个入口：

• import 路径 llama_index.core.workflow
• 也可独立 pip install llama-index-workflows

后台执行这一项是缺的。 官方框架文档未提供「关机续跑」的托管运行时， 进程要你自己跑。需要托管时走 LlamaAgents / llamactl：

"deploy them to LlamaCloud or your own infrastructure"

而那是平台侧。

There is no hosted execution that "keeps running after shutdown". But orchestration (Workflows) really is still in the OSS.

One spot is easy to misread from the README. The README lists Workflows under the LlamaParse platform's Agents section, while the official framework docs are explicit:

"llama-index-core comes with a stable version of Workflows included"

So orchestration has not been moved out of the OSS. Two entry points:

• Import path llama_index.core.workflow
• Also standalone: pip install llama-index-workflows

Background execution is the missing item. The official framework docs offer no hosted runtime for "keep-running-after-shutdown"; you run the process yourself. For hosting, go through LlamaAgents / llamactl:

"deploy them to LlamaCloud or your own infrastructure"

and that is the platform side.

工具与扩展Tools & extensions

工具层两件事：社区 ToolSpec，加一个官方的 MCP 接入包。

官方 tools 文档给了 FunctionTool。 还有 QueryEngineTool。

另有 "Community contributed ToolSpecs"。

处理「工具返回超大输出」的是 Utility Tools：

• OnDemandLoaderTool
• LoadAndSearchToolSpec

MCP 官方支持，而且是独立包。 文档原文："LlamaIndex provides robust support for consuming MCP servers through the llama-index-tools-mcp package"

它的 MCP 面：

• 三种传输都支持：SSE / Streamable HTTP / stdio
• 能把 Workflow 反发布为 MCP server
• 官方另托管 mcp.llamaindex.ai（文档检索 / LlamaParse 的 MCP 服务）

Two things at the tool layer: community ToolSpecs, plus one official MCP connector package.

The official tools docs give FunctionTool. There is also QueryEngineTool. Plus "Community contributed ToolSpecs".

Handling "a tool returns an oversized output" is what Utility Tools do:

• OnDemandLoaderTool
• LoadAndSearchToolSpec

MCP is officially supported, as a separate package. Docs wording: "LlamaIndex provides robust support for consuming MCP servers through the llama-index-tools-mcp package"

Its MCP surface:

• All three transports supported: SSE / Streamable HTTP / stdio
• A Workflow can be re-published as an MCP server
• The vendor also hosts mcp.llamaindex.ai (a docs-retrieval / LlamaParse MCP service)

上下文与记忆Context & memory

这一维度官方有具体机制，不是空白。

官方把框架叫 "data framework"。

核心是 structure your data（indices, graphs）。 再加 retrieval/query interface。

也就是「把外部数据变成可检索的上下文」。

检索结果怎么塞进上下文窗口？ 官方用 Response Synthesizer 回答：

• compact（默认）—— "stuff as many text ... that can fit within the context window"
• refine —— 逐块精炼
• tree_summarize —— 递归摘要
• simple_summarize —— 截断到单 prompt

这就是它的上下文裁剪与摘要机制：先按窗口装箱，装不下再切块，摘要类模式做压缩。

This dimension has concrete official mechanisms; it is not blank.

The official docs call the framework a "data framework".

The core is structure your data (indices, graphs). Plus a retrieval/query interface.

That is, "turn external data into retrievable context".

How do retrieval results get packed into the context window? The official answer is the Response Synthesizer:

• compact (default) — "stuff as many text ... that can fit within the context window"
• refine — refine chunk by chunk
• tree_summarize — recursive summarization
• simple_summarize — truncate to a single prompt

That is its context-trimming and summarization mechanism: pack to the window first, split into chunks when it does not fit, compress via the summarize-style modes.

权限与限制Permissions & limits

⚠ 这是一个明确的重大信息缺口，选型前必须自己确认。

官方 README 全文没写的东西：沙箱、文件权限模型、 网络访问控制、人工确认机制 —— 一个都没有。

与本站收录的其它编排框架对照：

• Google ADK 有 Tool Confirmation（HITL）
• CrewAI 有 guardrails（机制未说明）
• LangGraph 有 interrupts（可检查并修改状态）
• LlamaIndex —— README 全文里连一个名词都没有

⚠ 风险面的性质不一样。 它的主战场是文档解析与 RAG， 处理的通常是用户上传的数据。

这与「agent 动你本地文件」是两类风险（数据泄露面 vs 主机操作面）。

但 README 未讨论数据权限与隔离。

⚠ This is a clear, major information gap; confirm it yourself before choosing.

Things the official README never mentions: sandboxing, a file-permission model, network access controls, human-in-the-loop confirmation — not one.

Compared against the other orchestration frameworks this site files:

• Google ADK has Tool Confirmation (HITL)
• CrewAI has guardrails (mechanism undocumented)
• LangGraph has interrupts (state can be inspected and modified)
• LlamaIndex — the README does not even name the concept

⚠ The nature of the risk surface is different. Its main battleground is document parsing and RAG; what it handles is usually user-uploaded data.

That is a different class of risk from "an agent touching your local files" (data-leakage surface vs host-operation surface).

But the README does not discuss data permissions or isolation.

适合什么任务Best for

适合：把外部数据接进来做检索问答。

• 核心任务是把外部数据源（API / PDF / 文档 / SQL）接进来做检索与问答
• 需要 300+ 现成集成里的某一个
• 团队是 Python 栈
• 明确要用 LlamaParse 的解析能力（那走商业平台）

不太适合三种情况。

• 需要 agent 在自己项目目录里读写文件并跑命令 —— 那是 Deep Agents 那一类「工作区文件系统」
• 需要明确的沙箱与权限边界 —— 官方框架文档无此类机制
• 需要「关机后托管续跑」—— 走 LlamaAgents 平台侧

A fit: bringing external data in for retrieval-augmented Q&A.

• The core task is ingesting external data sources (APIs / PDFs / docs / SQL) for retrieval and Q&A
• You need one of the 300+ ready-made integrations
• The team is on the Python stack
• You explicitly want LlamaParse's parsing capabilities (that goes through the commercial platform)

Not a fit in three situations.

• You need an agent to read/write files in your own project directory and run commands — that is the Deep Agents class of "workspace file system"
• You need explicit sandboxing and permission boundaries — the official framework docs have no such mechanisms
• You need "hosted continuation after shutdown" — that goes through the LlamaAgents platform side

from llama_index.core.xxx import ClassABC      # 带 .core = 核心
from llama_index.xxx.yyy import SubclassABC    # 不带 = 集成包

from llama_index.core.xxx import ClassABC      # with .core = core
from llama_index.xxx.yyy import SubclassABC    # without = an integration package

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.