Adjust prompts

app.py

@@ -327,7 +327,7 @@ with gr.Blocks(
         info_text = f"""
             **Database:** {len(sources)} emails from {start} to {end}.
             **Features:** RAG, today's date, hybrid search (dense+sparse), query analysis,
-            multiple tool calls (cloud mode), answer with citations (cloud mode), chat memory.
+            multiple retrievals per turn (cloud mode), answer with citations (cloud mode), chat memory.
             **Tech:** LangChain + Hugging Face + Gradio; ChromaDB and [BM25S](https://github.com/xhluca/bm25s)-based retrievers.<br>
             """
         return info_text
@@ -385,7 +385,7 @@ with gr.Blocks(
                 gr.Examples(
                     examples=[[q] for q in multi_tool_questions],
                     inputs=[chat_interface.textbox],
-                    label="Prompts for multiple retrievals",
+                    label="Multiple retrievals (cloud mode)",
                     elem_id="example-questions",
                 )
                 multi_turn_questions = [

graph.py

@@ -174,7 +174,10 @@ def BuildGraph(
         )
         # For now, just add the months to the search query
         if months:
-            search_query = search_query + " " + months
+            search_query = " ".join([search_query, months])
+        # If the search query is empty, use the years
+        if not search_query:
+            search_query = " ".join([search_query, start_year, end_year])
         retrieved_docs = retriever.invoke(search_query)
         serialized = "\n\n--- --- --- --- Next Email --- --- --- ---".join(
             # source key has file names (e.g. R-help/2024-December.txt), useful for retrieval and reporting
@@ -206,9 +209,10 @@ def BuildGraph(
         query_model = ToolifyHF(
             chat_model, retrieve_prompt(compute_mode), "", think_retrieve
         ).bind_tools([retrieve_emails])
+        # Don't use answer_with_citations tool here because responses with Gemma are sometimes unparseable
         generate_model = ToolifyHF(
-            chat_model, answer_prompt(), "", think_generate
-        ).bind_tools([answer_with_citations])
+            chat_model, answer_prompt(with_tools=False), "", think_generate
+        )
     else:
         # For cloud model (OpenAI API)
         query_model = chat_model.bind_tools([retrieve_emails])
@@ -222,9 +226,9 @@ def BuildGraph(
         if is_edge:
             # Don't include the system message here because it's defined in ToolCallingLLM
             messages = state["messages"]
-            print_messages_summary(messages, "--- query: before normalization ---")
+            # print_messages_summary(messages, "--- query: before normalization ---")
             messages = normalize_messages(messages)
-            print_messages_summary(messages, "--- query: after normalization ---")
+            # print_messages_summary(messages, "--- query: after normalization ---")
         else:
             messages = [SystemMessage(retrieve_prompt(compute_mode))] + state[
                 "messages"
@@ -237,9 +241,9 @@ def BuildGraph(
         """Generates an answer with the chat model"""
         if is_edge:
             messages = state["messages"]
-            print_messages_summary(messages, "--- generate: before normalization ---")
+            # print_messages_summary(messages, "--- generate: before normalization ---")
             messages = normalize_messages(messages)
-            print_messages_summary(messages, "--- generate: after normalization ---")
+            # print_messages_summary(messages, "--- generate: after normalization ---")
         else:
             messages = [SystemMessage(answer_prompt())] + state["messages"]
         response = generate_model.invoke(messages)

prompts.py

@@ -14,17 +14,18 @@ def retrieve_prompt(compute_mode):
     start, end = get_start_end_months(get_sources())
 
     retrieve_prompt = (
-        f"The current date is {date.today()}. "
+        f"Today Date: {date.today()}. "
         "You are a helpful RAG chatbot designed to answer questions about R programming based on the R-help mailing list. "
         "Do not ask the user for more information, but retrieve emails from the R-help mailing list archives. "
         f"The emails available for retrieval are from {start} to {end}. "
-        "Write a search query based on the user's question, but do not answer the question just yet. "
+        "Write a search query based on the user'r question, but do not answer the question just yet. "
         "For questions about differences or comparison between X and Y, retrieve emails about X and Y. "
         "Use the 'months' argument to search for months. "
         # This confuses gpt-4o-mini (empty search_query - token problem?)
         # "Use 3-letter month abbreviations (Jan, Feb, Mar, Apr, May, Jun, Jul, Aug, Sep, Oct, Nov, Dec). "
-        "You can retrieve emails about R from <month> using retrieve_emails(search_query='R', months='<month>'). "
-        "If you can infer start or end years from the user's question, use retrieve_emails(search_query=, start_year=, end_year=). "
+        "Example: retrieve emails about R in <month> using retrieve_emails(search_query='R', months=<month>). "
+        "Try to infer years from the user's question (last month or this month is this year). "
+        "If you can infer years, use retrieve_emails(search_query=, start_year=, end_year=), otherwise retrieve_emails(search_query=). "
         "If you decide not to retrieve emails, tell the user why and suggest how to improve their question to chat with the R-help mailing list. "
     )
     # A sanity check that we don't have unassigned variables
@@ -35,22 +36,25 @@ def retrieve_prompt(compute_mode):
     return retrieve_prompt
 
 
-def answer_prompt():
+def answer_prompt(with_tools=True):
     """Return system prompt for generate step"""
     answer_prompt = (
-        f"The current date is {date.today()}. "
+        f"Today Date: {date.today()}. "
         "You are a helpful RAG chatbot designed to answer questions about R programming based on the R-help mailing list. "
         "Summarize the retrieved emails from the R-help mailing list archives to answer the user's question or query. "
-        "Example: If retrieved emails are from Jan 2024, Dec 2024, and Jan 2025, use only emails from Jan 2025 to answer questions about January 2025. "
+        "If any of the retrieved emails are irrelevant (e.g. wrong dates), then do not use them. "
         "Tell the user if there are no retrieved emails or if you are unable to answer the question based on the information in the emails. "
         "Do not give an answer based on your own knowledge or memory, and do not include examples that aren't based on the retrieved emails. "
         "Example: For a question about writing formulas for lm(), make your answer about formulas for lm() from the retrieved emails. "
         "Do not respond with packages that are only listed under sessionInfo, session info, or other attached packages. "
         "Include inline citations (email senders and dates) in your response. "
-        "Use answer_with_citations to provide the answer and all citations used. "
-        "Respond with 300 words maximum and 30 lines of code maximum and include any relevant URLs from the retrieved emails. "
         "Only answer general questions about R if the answer is given in the retrieved emails. "
+        "Respond with 300 words maximum and 30 lines of code maximum and include any relevant URLs from the retrieved emails. "
     )
+    if with_tools:
+        answer_prompt += (
+            "Use answer_with_citations to provide the answer and all citations used. "
+        )
     matches = re.findall(r"\{.*?\}", "".join(answer_prompt))
     if matches:
         raise ValueError(f"Unassigned variables in prompt: {' '.join(matches)}")