https://github.com/dzdx/learn-llama-index
https://github.com/dzdx/learn-llama-index
learning-tutorial llama-index
Last synced: 2 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/dzdx/learn-llama-index
- Owner: dzdx
- Created: 2023-11-22T15:23:33.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-04-28T12:07:43.000Z (about 1 year ago)
- Last Synced: 2025-03-25T09:53:00.014Z (3 months ago)
- Topics: learning-tutorial, llama-index
- Language: Python
- Homepage:
- Size: 12.2 MB
- Stars: 6
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 手把手入门 llama index
## 介绍
本项目可以搭建一个简单的问答系统，可以查询 北京、上海市、深圳市 在wiki上的信息，并进行简单的意图识别，进行简单闲聊.
主要用来演示如何用 llama index做一个简单的 RAG 应用
`query_todo` 和 `build_todo` 中的 TODO 为需要编写代码的部分，需要跑通 test.py 中所有的测试用例，也可以参考 `query`
和 `build` 两个模块代码的实现
## 运行准备
系统依赖
- \>=Python3.8
安装依赖
```bash
virtualenv venv
source venv/bin/activate
pip install -r requirements.txt
```
配置OpenAI ApiKey
```bash
export OPENAI_API_KEY=xxxx
```
## 补全 TODO
### 1. 下载 wiki 并进行解析
`pytest test.py::test_build_nodes`
```python
def build_nodes(data_file: str) -> List[BaseNode]:
# 把输入的文档解析为Document对象，txt的文件，一个文件一个Document
# 如果是pdf文件，会有特殊处理，一页一个Document
documents = SimpleDirectoryReader(input_files=[data_file]).load_data()
for doc in documents:
# 过滤掉自动生成的 `file_path` 这个metadata, 防止对embedding结果造成干扰
doc.excluded_llm_metadata_keys.append("file_path")
doc.excluded_embed_metadata_keys.append("file_path")
# 把 document 按句子进行分割成多个 nodes
return service_context.node_parser.get_nodes_from_documents(documents)
```
### 2. 构建单个城市的召回索引
`pytest test.py::test_build_index`
```python
def build_index(index_file: str, data_file: str):
if os.path.exists(index_file):
return
nodes = build_nodes(data_file)
# 两个index共用一个存储目录，可以复用DocumentStore
storage_context = StorageContext.from_defaults()
vector_index = VectorStoreIndex(nodes,
service_context=service_context,
storage_context=storage_context,
show_progress=True)
# TreeIndex的 num_children 是自底向上逐步summary, 生成parent node的时候，每个parent包含多少个children nodes
# summary_template 可以替换为中文的prompt，更稳定的得到中文的summary
tree_index = TreeIndex(nodes, num_children=8,
service_context=service_context,
storage_context=storage_context,
summary_template=CH_SUMMARY_PROMPT,
show_progress=True)
# 把两个索引的生成数据存储到index_file这个目录
storage_context.persist(persist_dir=index_file)
```
### 3. 构建三个城市的召回索引
`pytest test.py::test_build_all_index`
```python
def build_all():
# 单独构建全部的索引，后续查询会用到多索引
titles = ['北京市', '上海市', '深圳市']
for title in titles:
download_and_build_index(title, data_dir, index_dir)
```
### 4. 总结生成答案
`pytest test.py::test_response_synthesizer`
```python
def create_response_synthesizer(service_context: ServiceContext = None) -> BaseSynthesizer:
# 采用TreeSummarize的方式对多个上下文进行逐步总结，防止超过llm的context limit
# 同时用中文prompt得到更加稳定的中文summary
return get_response_synthesizer(
response_mode=ResponseMode.TREE_SUMMARIZE,
summary_template=CH_TREE_SUMMARIZE_PROMPT,
service_context=service_context,
)
```
### 5. 创建召回器
`pytest test.py::test_retriever`
```python
class DocumentQueryEngineFactory:
...
def create_retrievers(self):
ret = []
for index in self.indices:
# 每个索引可能会用不同的生成retriever的方式
# 取 `similarity_top_k` 和query 最相似的 node
if isinstance(index, VectorStoreIndex):
ret.append(index.as_retriever(similarity_top_k=8))
if isinstance(index, TreeIndex):
# TreeRetriever采用自顶向下逐步检索的方式得到叶子节点
ret.append(index.as_retriever(retriever_mode=TreeRetrieverMode.SELECT_LEAF_EMBEDDING))
return ret
```
### 6. 创建单个文档的查询引擎
`pytest test.py::test_query_engine`
```python
def create_query_engine(self, service_context: ServiceContext) -> RetrieverQueryEngine:
# 组合多索引召回、排序，答案合成，构建最终的query engine
retriever = MultiRetriever(self.create_retrievers())
# LLMRerank只选取最相关的top_n, 进一步提高命中率，防止召回阶段拿到不相关的内容
node_postprocessors = [
LLMRerank(top_n=4, choice_batch_size=2, choice_select_prompt=CH_CHOICE_SELECT_PROMPT,
service_context=service_context)
]
return RetrieverQueryEngine.from_args(
retriever,
node_postprocessors=node_postprocessors,
service_context=service_context,
response_synthesizer=create_response_synthesizer(service_context)
)
```
### 7. 创建组合查询引擎
`pytest test.py::test_compose_query_engine`
```python
def create_compose_query_engine(city_indices: Dict[str, List[BaseIndex]],
service_context: ServiceContext) -> BaseQueryEngine:
query_engines = []
for city, indices in city_indices.items():
summary = f"""
此内容包含关于{city}的维基百科文章。
如果您需要查找有关{city}的具体事实，请使用此索引。"
如果您想分析多个城市，请不要使用此索引。
"""
query_engines.append(DocumentQueryEngineFactory(
indices=indices,
summary=summary
))
# 为每个城市的query engine指定一个summary，ComposableGraph会使用大模型来判断问题需要使用哪一个QueryEngine
graph = ComposableGraph.from_indices(
TreeIndex,
[e.first_index() for e in query_engines],
[e.summary for e in query_engines],
service_context=service_context,
)
return graph.as_query_engine(
response_synthesizer=create_response_synthesizer(service_context=service_context),
custom_query_engines={e.first_index().index_id: e.create_query_engine(service_context) for e in query_engines},
service_context=service_context,
query_template=CH_QUERY_PROMPT,
)
```
### 9. 简单的路由分发
`pytest test.py::test_simple_route`
```python
def create_route_query_engine(query_engines: List[BaseQueryEngine], descriptions: List[str],
service_context: ServiceContext = None):
assert len(query_engines) == len(descriptions)
tools = []
for i, query_engine in enumerate(query_engines):
query_tool = QueryEngineTool.from_defaults(
query_engine=query_engine,
description=descriptions[i]
)
tools.append(query_tool)
return llama_index.query_engine.RouterQueryEngine(
selector=LLMSingleSelector.from_defaults(service_context=service_context,
prompt_template_str=CH_SINGLE_SELECT_PROMPT_TMPL),
service_context=service_context,
query_engine_tools=tools
)
```
### 10. 意图识别
`pytest test.py::test_route_query_engine`
```python
class Chatter:
...
def create_query_engine(self):
index_query_engine = create_compose_query_engine(self.city_indices, self.service_context)
index_summary = f"提供 {', '.join(self.city_indices.keys())} 这几个城市的相关信息"
llm_query_engine = LlmQueryEngine(llm=self.llm, callback_manager=self.cb_manager)
llm_summary = "提供其他所有信息"
route_query_engine = create_route_query_engine(
[index_query_engine, llm_query_engine],
[index_summary, llm_summary],
service_context=self.service_context)
return route_query_engine
```
# 完整运行项目
```bash
python main.py
Enter a query:北京气候如何
```