feat: using a detailed prompt for the VLM.

src/insurance_assistants/agents.py

@@ -16,7 +16,7 @@ from smolagents import (
 )
 
 from src.insurance_assistants.complex_rag import RAG
-from src.insurance_assistants.consts import PROMPT_PREFIX
+from src.insurance_assistants.consts import PROMPT_PREFIX, PROMPT
 
 # _ = load_dotenv(dotenv_path=find_dotenv())
 rag_app = RAG()
@@ -45,9 +45,11 @@ class InsuranceInfoRetriever(Tool):
         chat_template = [
             {
                 "role": "system",
-                "content": """You find answers from the relevant documents. Answer only 
-        from these documents. If answer isn't available return 'Question cannot be answered based
-        on the documents provided.' """,
+                "content": PROMPT
+        #         """You find answers from the relevant documents. Answer only 
+        # from these documents. If answer isn't available return 'Question cannot be answered based
+        # on the documents provided.' """
+                ,
             },
             {
                 "role": "user",
@@ -76,7 +78,8 @@ class InsuranceInfoRetriever(Tool):
 insurance_agent = CodeAgent(
     tools=[InsuranceInfoRetriever(), FinalAnswerTool()],
     model=InferenceClientModel(bill_to="VitalNest", temperature=0.1,
-                              token=os.getenv("HF_TOKEN")),
+                              # token=os.getenv("HF_TOKEN")
+                              ),
     additional_authorized_imports=["os", "requests", "bs4", "pil", "base64", "io"],
     max_steps=1,
     verbosity_level=-1,

src/insurance_assistants/consts.py

@@ -10,11 +10,9 @@ Give detailed and extensive answers, only containing info in the pages you are g
 You can answer using information contained in plots and figures if necessary.
 Answer in the same language as the query.
 
-Query: {query}
-PDF pages:
 """
 
-PRIMARY_HEADING = """# An Agentic RAG for Health Insurance Documents
+PRIMARY_HEADING = """# Multi-agent Health Insurance Advisor
 #### This agent answers health insurance related questions from pre-ingested set \
 of health insurance documents, search wikipedia, search the web \
 and execute basic python code. \