better tutoring flow

agent.py

@@ -17,6 +17,11 @@ from langchain.memory import ConversationBufferWindowMemory
 import random
 from pydantic import Extra
 
+from langchain.chains import LLMChain
+from langchain import PromptTemplate
+
+
+
 
 import promptlayer
 from langchain.callbacks import PromptLayerCallbackHandler
@@ -67,9 +72,67 @@ class GetrandomTool(BaseTool):
         
         return text
 
-    def _arun(self, value: Union[int, float]):
+    def _arun(self, question: str):
+        raise NotImplementedError("This tool does not support async")
+
+
+#Tutor tool?
+#A chain with retrieval for answering, and a constant input summary of the
+#tutoring flow so far
+
+class TutoringTool(BaseTool):
+    name = "TutoringTool"
+
+    description = """This tool is capable of generating tutoring questions. 
+    It has to be called with a summary of the previous tutoring discussion steps,
+    or in case of a new tutoring session, with a randomly chosen piece of material.
+    As for it's output, it has to be kept at it is, sent bakc to the user."""
+
+    class Config:
+        extra = Extra.allow
+
+    def _run(self, question: str):
+        # initiate chain and prompts
+        # Would be waaay more elegant to have it at init time, but...
+        prompt_template = """
+        You act as a knowledgeable tutor. Based on some previous [apropos] 
+        (a question, a piace of material, or a summary of a tutoring session) 
+        and some [relevant documents] generate a tutoring question, 
+        that helps in systematically think about the topic at hand and 
+        for which the answer is deepening the knowledge of the subject, getting closser to an aswer.
+        You should NOT answer the question at hand, just either ask a helping question 
+        or confirm if an aswer is correct! This should be [your output].
+
+        [apropos]
+        {apropos}
+
+        [relevant_documents]
+        {relevant_documents}
+
+        [your output]:
+        """
+
+        prompt = PromptTemplate(
+            input_variables=["apropos", "relevant_documents"],
+            template=prompt_template,
+        )
+
+        # do retrieval
+        relevant_documents = self.retriever.get_relevant_documents(question)
+
+        # concat the two
+        # execute a chain
+        llm = ChatOpenAI(model_name=os.environ["CHAT_MODEL"])
+        chain = LLMChain(llm=llm, prompt=prompt)
+
+        result = chain.run(apropos=question, relevant_documents=relevant_documents)
+        return result
+    
+    def _arun(self, question: str):
         raise NotImplementedError("This tool does not support async")
 
+
+
 class QMLAgent():
 
     def __init__(self):
@@ -98,6 +161,8 @@ class QMLAgent():
         random_tool.indexer = index
         random_tool.index_max = index.describe_index_stats()["total_vector_count"] 
 
+        tutoring_tool = TutoringTool()
+        tutoring_tool.retriever = retriever
 
         tools = [
             Tool(
@@ -119,6 +184,12 @@ class QMLAgent():
                 #return_direct=False
 
             ),
+            Tool.from_function(
+                name="Tutoring",
+                func=tutoring_tool._run,
+                description=tutoring_tool.description,
+                return_direct=True
+            ),
         ]
 
         memory = ConversationBufferWindowMemory(k=os.environ["MEMORY_LENGTH"], memory_key="chat_history", return_messages=True)