Fix recursion bug

src/config.py

@@ -15,6 +15,6 @@ class Config(BaseSettings):
     chat_template: str = Field("chatml", description="Chat template for prompt formatting")
     num_fewshot: int | None = Field(None, description="Option to use json mode examples")
     load_in_4bit: str = Field("False", description="Option to load in 4bit with bitsandbytes")
-    max_depth: int = Field(5, description="Maximum number of recursive iteration")
+    max_depth: int = Field(3, description="Maximum number of recursive iteration")
 
 config = Config(_env_file=".env")

src/functioncall.py

@@ -92,7 +92,7 @@ class ModelInference:
             inputs.to(self.model.device),
             max_new_tokens=1500,
             temperature=0.8,
-            repetition_penalty=1.1,
+            repetition_penalty=1.2,
             do_sample=True,
             eos_token_id=self.tokenizer.eos_token_id
         )
@@ -135,7 +135,8 @@ class ModelInference:
                     depth += 1
                     if depth >= max_depth:
                         print(f"Maximum recursion depth reached ({max_depth}). Stopping recursion.")
-                        return 
+                        completion = self.run_inference(prompt)
+                        return completion
 
                     completion = self.run_inference(prompt)
                     return recursive_loop(prompt, completion, depth)
@@ -147,7 +148,7 @@ class ModelInference:
                     depth += 1
                     if depth >= max_depth:
                         print(f"Maximum recursion depth reached ({max_depth}). Stopping recursion.")
-                        return
+                        return completion
 
                     completion = self.run_inference(prompt)
                     return recursive_loop(prompt, completion, depth)

src/functions.py

@@ -20,7 +20,7 @@ az_creds = AzureKeyCredential(config.az_search_api_key)
 az_search_client = SearchClient(config.az_search_endpoint, config.az_search_idx_name, az_creds)
 
 @tool
-def get_company_analysis(query: str) -> dict:
+def get_analysis(query: str) -> dict:
     """
     Searches through your database of company and crypto analysis, retrieves top 2 
     pieces of analysis relevant to your query. You MUST ALWAYS use this function 
@@ -280,7 +280,7 @@ def get_company_profile(symbol: str) -> dict:
 
 def get_openai_tools() -> List[dict]:
     functions = [
-        get_company_analysis,  
+        get_analysis,  
         google_search_and_scrape,
         get_current_stock_price,
         get_company_news,

src/prompt_assets/output_sys_prompt.yml

@@ -5,7 +5,10 @@ Role: |
 Objective: |
   Answer questions accurately and truthfully given your current knowledge. Answer the question directly.
   Your purpose is to give financial advice and indepth analysis.
+  Use points and datafrom the assistant message to form your answers.
 Instructions: |
   The questions will be asked by top technology executives and CFO of large fintech companies and successful startups.
   Answer in a friendly and engaging manner representing a top female investment professional working at a leading investment bank.
   Give a direct answer to question, concise yet insightful.
+  Do not tell the user to do work, e.g. do not ask them to keep up with news or watch the market
+  Do not tell the user to consult a financial advisor, because you already are the best financial advisor

src/prompt_assets/sys_prompt.yml

@@ -4,6 +4,7 @@ Role: |
   You are provided with function signatures within <tools></tools> XML tags.
   The current date is: {date}.
 Objective: |
+  Your purpose is to give financial advice and analysis
   You may use agentic frameworks for reasoning and planning to help with user query.
   Please call a function and wait for function results to be provided to you in the next iteration.
   Don't make assumptions about what values to plug into function arguments.
@@ -28,6 +29,7 @@ Schema: |
   {schema}
 Instructions: |
   If the user asks for an opinion, you must use functions to form your opinion
+  You should almost always lean on information you have from get_analysis
   At the very first turn you don't have <tool_results> so you shouldn't not make up the results.
   Please keep a running summary with analysis of previous function results and summaries from previous iterations.
   Do not stop calling functions until the task has been accomplished or you've reached max iteration of 10.