[RetrievalAgent] Add base retrieval agent #258
Conversation
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## develop #258 +/- ##
===========================================
+ Coverage 69.37% 70.09% +0.71%
===========================================
Files 63 64 +1
Lines 3236 3347 +111
===========================================
+ Hits 2245 2346 +101
- Misses 991 1001 +10 ☔ View full report in Codecov by Sentry. |
| QUERY_DECOMPOSITION = """请把下面的问题分解成子问题，每个子问题必须足够简单，要求： | ||
| 1.严格按照【JSON格式】的形式输出：{'子问题1':'具体子问题1','子问题2':'具体子问题2'} | ||
| 问题：{{prompt}} 子问题：""" |
There was a problem hiding this comment.
加了一个few shot retriever，更通用
| self.use_extractor = use_extractor | ||
| self.extractor = PromptTemplate(CONTENT_COMPRESSOR, input_variables=["context", "query"]) | ||
|
|
||
| async def _run(self, prompt: str, files: Optional[List[File]] = None) -> AgentResponse: |
There was a problem hiding this comment.
- 为啥不写在_run里面呢
- log需要加，同时steps需要遵守， 要不然返回最后的response信息不足
There was a problem hiding this comment.
run_llm里面加了logger日志
| else: | ||
| steps_input = HumanMessage(content=self.query_transform.format(query=prompt)) |
There was a problem hiding this comment.
zero shot，完全靠大模型自己的能力进行子query分解的时候
| """ | ||
|
|
||
|
|
||
| class FaissFewShotSearch: |
There was a problem hiding this comment.
FaissFewShotSearch和FaissAbstractSearch 这2个类都以Faiss打头，但是没看出和faiss有任何关系. 唯一有关系的是self.db. similarity_search_with_relevance_scores这个函数, 但是这个函数应该是任何vector db实现都有的，也不仅仅和Faiss相关
There was a problem hiding this comment.
已修改，这个similarity_search_with_relevance_scores函数是langchain独有的
| @dataclass | ||
| class RetrievalStep(AgentStep): | ||
| """A step taken by an agent.""" | ||
|
|
||
| name: str |
There was a problem hiding this comment.
如果只是为了加一个name字段，就没有必要去新开一个class了。直接将name字段放入info就可以了，info是一个dict
There was a problem hiding this comment.
已修改，但我有个疑问：是不是可以给每个step加一个别名，当成默认属性，然后单独拿出来？
w5688414
changed the title
|
|
||
|
|
||
| class LangChainRetrievalTool(Tool): | ||
| description: str = "在知识库中检索与用户输入query相关的段落" |
There was a problem hiding this comment.
你这里没有设置 InputView 和 OutputView。
There was a problem hiding this comment.
langchain里面有metadata字段，是一个Dict，用于存储元数据，不太好实例化成pydantic的形式
TODO
BFS方法 (Base版)
Graph DFS方法（高级版）