Merge pull request #1721 from siiddhantt/feat/react-agent

feat: ReActAgent and agent refactor

Author: dartpain

application/agents/__init__.py

@@ -1,9 +1,11 @@
 from application.agents.classic_agent import ClassicAgent
+from application.agents.react_agent import ReActAgent
 
 
 class AgentCreator:
     agents = {
         "classic": ClassicAgent,
+        "react": ReActAgent,
     }
 
     @classmethod

application/agents/agent_creator.py

@@ -1,38 +1,60 @@
-from typing import Dict, Generator
+import uuid
+from abc import ABC, abstractmethod
+from typing import Dict, Generator, List, Optional
 
 from application.agents.llm_handler import get_llm_handler
 from application.agents.tools.tool_action_parser import ToolActionParser
 from application.agents.tools.tool_manager import ToolManager
 
 from application.core.mongo_db import MongoDB
 from application.llm.llm_creator import LLMCreator
+from application.logging import build_stack_data, log_activity, LogContext
+from application.retriever.base import BaseRetriever
 
 
-class BaseAgent:
+class BaseAgent(ABC):
     def __init__(
         self,
-        endpoint,
-        llm_name,
-        gpt_model,
-        api_key,
-        user_api_key=None,
-        decoded_token=None,
+        endpoint: str,
+        llm_name: str,
+        gpt_model: str,
+        api_key: str,
+        user_api_key: Optional[str] = None,
+        prompt: str = "",
+        chat_history: Optional[List[Dict]] = None,
+        decoded_token: Optional[Dict] = None,
     ):
         self.endpoint = endpoint
+        self.llm_name = llm_name
+        self.gpt_model = gpt_model
+        self.api_key = api_key
+        self.user_api_key = user_api_key
+        self.prompt = prompt
+        self.decoded_token = decoded_token or {}
+        self.user: str = decoded_token.get("sub")
+        self.tool_config: Dict = {}
+        self.tools: List[Dict] = []
+        self.tool_calls: List[Dict] = []
+        self.chat_history: List[Dict] = chat_history if chat_history is not None else []
         self.llm = LLMCreator.create_llm(
             llm_name,
             api_key=api_key,
             user_api_key=user_api_key,
             decoded_token=decoded_token,
         )
         self.llm_handler = get_llm_handler(llm_name)
-        self.gpt_model = gpt_model
-        self.tools = []
-        self.tool_config = {}
-        self.tool_calls = []
 
-    def gen(self, *args, **kwargs) -> Generator[Dict, None, None]:
-        raise NotImplementedError('Method "gen" must be implemented in the child class')
+    @log_activity()
+    def gen(
+        self, query: str, retriever: BaseRetriever, log_context: LogContext = None
+    ) -> Generator[Dict, None, None]:
+        yield from self._gen_inner(query, retriever, log_context)
+
+    @abstractmethod
+    def _gen_inner(
+        self, query: str, retriever: BaseRetriever, log_context: LogContext
+    ) -> Generator[Dict, None, None]:
+        pass
 
     def _get_user_tools(self, user="local"):
         mongo = MongoDB.get_client()
@@ -109,14 +131,12 @@ def _execute_tool_action(self, tools_dict, call):
                 for param, details in action_data[param_type]["properties"].items():
                     if param not in call_args and "value" in details:
                         target_dict[param] = details["value"]
-
         for param, value in call_args.items():
             for param_type, target_dict in param_types.items():
                 if param_type in action_data and param in action_data[param_type].get(
                     "properties", {}
                 ):
                     target_dict[param] = value
-
         tm = ToolManager(config={})
         tool = tm.load_tool(
             tool_data["name"],
@@ -151,3 +171,79 @@ def _execute_tool_action(self, tools_dict, call):
         self.tool_calls.append(tool_call_data)
 
         return result, call_id
+
+    def _build_messages(
+        self,
+        system_prompt: str,
+        query: str,
+        retrieved_data: List[Dict],
+    ) -> List[Dict]:
+        docs_together = "\n".join([doc["text"] for doc in retrieved_data])
+        p_chat_combine = system_prompt.replace("{summaries}", docs_together)
+        messages_combine = [{"role": "system", "content": p_chat_combine}]
+
+        for i in self.chat_history:
+            if "prompt" in i and "response" in i:
+                messages_combine.append({"role": "user", "content": i["prompt"]})
+                messages_combine.append({"role": "assistant", "content": i["response"]})
+            if "tool_calls" in i:
+                for tool_call in i["tool_calls"]:
+                    call_id = tool_call.get("call_id") or str(uuid.uuid4())
+
+                    function_call_dict = {
+                        "function_call": {
+                            "name": tool_call.get("action_name"),
+                            "args": tool_call.get("arguments"),
+                            "call_id": call_id,
+                        }
+                    }
+                    function_response_dict = {
+                        "function_response": {
+                            "name": tool_call.get("action_name"),
+                            "response": {"result": tool_call.get("result")},
+                            "call_id": call_id,
+                        }
+                    }
+
+                    messages_combine.append(
+                        {"role": "assistant", "content": [function_call_dict]}
+                    )
+                    messages_combine.append(
+                        {"role": "tool", "content": [function_response_dict]}
+                    )
+        messages_combine.append({"role": "user", "content": query})
+        return messages_combine
+
+    def _retriever_search(
+        self,
+        retriever: BaseRetriever,
+        query: str,
+        log_context: Optional[LogContext] = None,
+    ) -> List[Dict]:
+        retrieved_data = retriever.search(query)
+        if log_context:
+            data = build_stack_data(retriever, exclude_attributes=["llm"])
+            log_context.stacks.append({"component": "retriever", "data": data})
+        return retrieved_data
+
+    def _llm_gen(self, messages: List[Dict], log_context: Optional[LogContext] = None):
+        resp = self.llm.gen_stream(
+            model=self.gpt_model, messages=messages, tools=self.tools
+        )
+        if log_context:
+            data = build_stack_data(self.llm)
+            log_context.stacks.append({"component": "llm", "data": data})
+        return resp
+
+    def _llm_handler(
+        self,
+        resp,
+        tools_dict: Dict,
+        messages: List[Dict],
+        log_context: Optional[LogContext] = None,
+    ):
+        resp = self.llm_handler.handle_response(self, resp, tools_dict, messages)
+        if log_context:
+            data = build_stack_data(self.llm_handler)
+            log_context.stacks.append({"component": "llm_handler", "data": data})
+        return resp

application/agents/base.py

@@ -1,86 +1,23 @@
-import uuid
 from typing import Dict, Generator
 
 from application.agents.base import BaseAgent
-from application.logging import build_stack_data, log_activity, LogContext
+from application.logging import LogContext
 
 from application.retriever.base import BaseRetriever
 
 
 class ClassicAgent(BaseAgent):
-    def __init__(
-        self,
-        endpoint,
-        llm_name,
-        gpt_model,
-        api_key,
-        user_api_key=None,
-        prompt="",
-        chat_history=None,
-        decoded_token=None,
-    ):
-        super().__init__(
-            endpoint, llm_name, gpt_model, api_key, user_api_key, decoded_token
-        )
-        self.user = decoded_token.get("sub")
-        self.prompt = prompt
-        self.chat_history = chat_history if chat_history is not None else []
-
-    @log_activity()
-    def gen(
-        self, query: str, retriever: BaseRetriever, log_context: LogContext = None
-    ) -> Generator[Dict, None, None]:
-        yield from self._gen_inner(query, retriever, log_context)
-
     def _gen_inner(
         self, query: str, retriever: BaseRetriever, log_context: LogContext
     ) -> Generator[Dict, None, None]:
         retrieved_data = self._retriever_search(retriever, query, log_context)
 
-        docs_together = "\n".join([doc["text"] for doc in retrieved_data])
-        p_chat_combine = self.prompt.replace("{summaries}", docs_together)
-        messages_combine = [{"role": "system", "content": p_chat_combine}]
-
-        if len(self.chat_history) > 0:
-            for i in self.chat_history:
-                if "prompt" in i and "response" in i:
-                    messages_combine.append({"role": "user", "content": i["prompt"]})
-                    messages_combine.append(
-                        {"role": "assistant", "content": i["response"]}
-                    )
-                if "tool_calls" in i:
-                    for tool_call in i["tool_calls"]:
-                        call_id = tool_call.get("call_id")
-                        if call_id is None or call_id == "None":
-                            call_id = str(uuid.uuid4())
-
-                        function_call_dict = {
-                            "function_call": {
-                                "name": tool_call.get("action_name"),
-                                "args": tool_call.get("arguments"),
-                                "call_id": call_id,
-                            }
-                        }
-                        function_response_dict = {
-                            "function_response": {
-                                "name": tool_call.get("action_name"),
-                                "response": {"result": tool_call.get("result")},
-                                "call_id": call_id,
-                            }
-                        }
-
-                        messages_combine.append(
-                            {"role": "assistant", "content": [function_call_dict]}
-                        )
-                        messages_combine.append(
-                            {"role": "tool", "content": [function_response_dict]}
-                        )
-        messages_combine.append({"role": "user", "content": query})
-
         tools_dict = self._get_user_tools(self.user)
         self._prepare_tools(tools_dict)
 
-        resp = self._llm_gen(messages_combine, log_context)
+        messages = self._build_messages(self.prompt, query, retrieved_data)
+
+        resp = self._llm_gen(messages, log_context)
 
         if isinstance(resp, str):
             yield {"answer": resp}
@@ -93,7 +30,7 @@ def _gen_inner(
             yield {"answer": resp.message.content}
             return
 
-        resp = self._llm_handler(resp, tools_dict, messages_combine, log_context)
+        resp = self._llm_handler(resp, tools_dict, messages, log_context)
 
         if isinstance(resp, str):
             yield {"answer": resp}
@@ -105,36 +42,11 @@ def _gen_inner(
             yield {"answer": resp.message.content}
         else:
             completion = self.llm.gen_stream(
-                model=self.gpt_model, messages=messages_combine, tools=self.tools
+                model=self.gpt_model, messages=messages, tools=self.tools
             )
             for line in completion:
                 if isinstance(line, str):
                     yield {"answer": line}
 
         yield {"sources": retrieved_data}
         yield {"tool_calls": self.tool_calls.copy()}
-
-    def _retriever_search(self, retriever, query, log_context):
-        retrieved_data = retriever.search(query)
-        if log_context:
-            data = build_stack_data(retriever, exclude_attributes=["llm"])
-            log_context.stacks.append({"component": "retriever", "data": data})
-        return retrieved_data
-
-    def _llm_gen(self, messages_combine, log_context):
-        resp = self.llm.gen_stream(
-            model=self.gpt_model, messages=messages_combine, tools=self.tools
-        )
-        if log_context:
-            data = build_stack_data(self.llm)
-            log_context.stacks.append({"component": "llm", "data": data})
-        return resp
-
-    def _llm_handler(self, resp, tools_dict, messages_combine, log_context):
-        resp = self.llm_handler.handle_response(
-            self, resp, tools_dict, messages_combine
-        )
-        if log_context:
-            data = build_stack_data(self.llm_handler)
-            log_context.stacks.append({"component": "llm_handler", "data": data})
-        return resp