update retriever

.gitignore

@@ -19,6 +19,7 @@ log
 logs
 *.log.*
 *.db
+*.lock
 
 ### IDE
 .vscode

src/api.py

@@ -50,7 +50,7 @@ def chat():
     if config.enable_knowledge_base:
         kb_res = pre_retrieval.search(query)
         if kb_res:
-            kb_res = "\n".join([f"{r['id']}: {r['entity']['text']}" for r in kb_res[0]])
+            kb_res = "\n".join([f"{r['id']}: {r['entity']['text']}" for r in kb_res])
             kb_res = f"知识库信息: {kb_res}"
             external += kb_res
 

src/cli.py

@@ -1,7 +1,7 @@
 import os
 from dotenv import load_dotenv
 from core import HistoryManager
-from core import PreRetrieval
+from core import PreRetrieval, Retriever
 from config import Config
 from models import select_model
 
@@ -11,30 +11,23 @@ load_dotenv()
 if __name__ == "__main__":
     config = Config("config/base.yaml")
     model = select_model(config)
-    pre_retrieval = PreRetrieval(config)
+    retriever = Retriever(config)
     # pre_retrieval.add_file("/home/zwj/workspace/ProjectAthena/src/data/file/鉴定工作报告、技术报告-0708.pdf")
 
     print(f"[{config.model_provider}:{config.get('model_name', 'default')}] Type 'exit' to quit")
 
     history_manager = HistoryManager()
     while True:
-        message = input("\nUser: ")
-        if message == "exit":
+        query = input("\nUser: ")
+        if query == "exit":
             break
 
-        external = ""
+        # 检索结果
+        refs = retriever(query)
+        # 重新构建用户的输入
+        query = retriever.construct_query(query, refs)
 
-        if config.enable_knowledge_base:
-            kb_res = pre_retrieval.search(message)
-            if kb_res:
-                kb_res = "\n".join([f"{r['id']}: {r['entity']['text']}" for r in kb_res[0]])
-                kb_res = f"知识库信息: {kb_res}"
-                external += kb_res
-
-        if len(external) > 0:
-            message = f"以下是参考资料：\n\n\n {external} 请根据前面的知识回答：{message}"
-
-        messages = history_manager.add_user(message)
+        messages = history_manager.add_user(query)
         response = model.predict(messages, stream=config.stream)
 
         if config.stream:

src/core/__init__.py

@@ -1,2 +1,3 @@
 from .history import *
-from .preretrieval import *
+from .preretrieval import *
+from .retriever import *

src/core/preretrieval.py

@@ -88,7 +88,7 @@ class PreRetrieval:
             output_fields=["text", "subject"],  # specifies fields to be returned
         )
 
-        return res
+        return res[0]  # 因为 query 只有一个
 
     def read_text(self, file):
         support_format = [".pdf", ".txt", "*.md"]

src/core/retriever.py

@@ -0,0 +1,50 @@
+from core import PreRetrieval
+
+class Retriever:
+
+    def __init__(self, config):
+        self.config = config
+        self.pre_retrieval = PreRetrieval(config)
+
+    def retrieval(self, query):
+
+        refs = {}
+
+        # TODO: 查询分类、查询重写、查询分解、查询伪文档生成（HyDE）
+
+        if self.config.enable_knowledge_base:
+            refs["knowledge_base"] = self.pre_retrieval.search(query)
+
+        return refs
+
+    def construct_query(self, query, refs):
+        # TODO：Reranking
+
+        if len(refs) == 0:
+            return query
+
+        external = ""
+
+        kb_res = refs.get("knowledge_base")
+        if kb_res:
+            kb_text = "\n".join([f"{r['id']}: {r['entity']['text']}" for r in kb_res])
+            external += f"知识库信息: \n\n{kb_text}"
+
+        if len(external) > 0:
+            query = f"以下是参考资料：\n\n\n{external}\n\n\n请根据前面的知识回答：{query}"
+
+        return query
+
+    def query_classification(self, query):
+        """判断是否需要查询
+        - 对于完全基于用户给定信息的任务，称之为“足够”“sufficient”，不需要检索；
+        - 否则，称之为“不足”“insufficient”，可能需要检索，
+        """
+        raise NotImplementedError
+
+    def rewrite_query(self, query):
+        """重写查询"""
+        raise NotImplementedError
+
+    def __call__(self, query):
+        return self.retrieval(query)