Extract prompts as LangChain Prompt Templates into a separate module

Improves code modularity, cleanliness. Reduces bloat in GPT.py module

pyproject.toml

@@ -53,6 +53,7 @@ dependencies = [
     "torch == 1.13.1",
     "uvicorn == 0.17.6",
     "aiohttp == 3.8.4",
+    "langchain >= 0.0.187",
 ]
 dynamic = ["version"]
 

src/khoj/processor/conversation/gpt.py

@@ -5,10 +5,10 @@ from datetime import datetime
 
 # Internal Packages
 from khoj.utils.constants import empty_escape_sequences
+from khoj.processor.conversation import prompts
 from khoj.processor.conversation.utils import (
     chat_completion_with_backoff,
     completion_with_backoff,
-    message_to_prompt,
     generate_chatml_messages_with_context,
 )
 
@@ -20,17 +20,9 @@ def answer(text, user_query, model, api_key=None, temperature=0.5, max_tokens=50
     """
     Answer user query using provided text as reference with OpenAI's GPT
     """
-    # Setup Prompt based on Summary Type
+    # Setup Prompt from arguments
-    prompt = f"""
+    prompt = prompts.answer.format(text=text, user_query=user_query)
-You are a friendly, helpful personal assistant.
-Using the users notes below, answer their following question. If the answer is not contained within the notes, say "I don't know."
 
-Notes:
-{text}
-
-Question: {user_query}
-
-Answer (in second person):"""
     # Get Response from GPT
     logger.debug(f"Prompt for GPT: {prompt}")
     response = completion_with_backoff(
@@ -53,19 +45,9 @@ def summarize(text, summary_type, model, user_query=None, api_key=None, temperat
     """
     # Setup Prompt based on Summary Type
     if summary_type == "chat":
-        prompt = f"""
+        prompt = prompts.summarize_chat.format(text=text)
-You are an AI. Summarize the conversation below from your perspective:
-
-{text}
-
-Summarize the conversation from the AI's first-person perspective:"""
     elif summary_type == "notes":
-        prompt = f"""
+        prompt = prompts.summarize_notes.format(text=text, user_query=user_query)
-Summarize the below notes about {user_query}:
-
-{text}
-
-Summarize the notes in second person perspective:"""
 
     # Get Response from GPT
     logger.debug(f"Prompt for GPT: {prompt}")
@@ -102,63 +84,16 @@ def extract_questions(text, model="text-davinci-003", conversation_log={}, api_k
     current_new_year = today.replace(month=1, day=1)
     last_new_year = current_new_year.replace(year=today.year - 1)
 
-    prompt = f"""
+    prompt = prompts.extract_questions.format(
-You are Khoj, an extremely smart and helpful search assistant with the ability to retrieve information from the users notes.
+        current_date=today.strftime("%A, %Y-%m-%d"),
-- The user will provide their questions and answers to you for context.
+        last_new_year=last_new_year.strftime("%Y"),
-- Add as much context from the previous questions and answers as required into your search queries.
+        last_new_year_date=last_new_year.strftime("%Y-%m-%d"),
-- Break messages into multiple search queries when required to retrieve the relevant information.
+        current_new_year_date=current_new_year.strftime("%Y-%m-%d"),
-- Add date filters to your search queries from questions and answers when required to retrieve the relevant information.
+        bob_tom_age_difference={current_new_year.year - 1984 - 30},
-
+        bob_age={current_new_year.year - 1984},
-What searches, if any, will you need to perform to answer the users question?
+        chat_history=chat_history,
-Provide search queries as a JSON list of strings
+        text=text,
-Current Date: {today.strftime("%A, %Y-%m-%d")}
+    )
-
-Q: How was my trip to Cambodia?
-
-["How was my trip to Cambodia?"]
-
-A: The trip was amazing. I went to the Angkor Wat temple and it was beautiful.
-
-Q: Who did i visit that temple with?
-
-["Who did I visit the Angkor Wat Temple in Cambodia with?"]
-
-A: You visited the Angkor Wat Temple in Cambodia with Pablo, Namita and Xi.
-
-Q: What national parks did I go to last year?
-
-["National park I visited in {last_new_year.strftime("%Y")} dt>=\\"{last_new_year.strftime("%Y-%m-%d")}\\" dt<\\"{current_new_year.strftime("%Y-%m-%d")}\\""]
-
-A: You visited the Grand Canyon and Yellowstone National Park in {last_new_year.strftime("%Y")}.
-
-Q: How are you feeling today?
-
-[]
-
-A: I'm feeling a little bored. Helping you will hopefully make me feel better!
-
-Q: How many tennis balls fit in the back of a 2002 Honda Civic?
-
-["What is the size of a tennis ball?", "What is the trunk size of a 2002 Honda Civic?"]
-
-A: 1085 tennis balls will fit in the trunk of a Honda Civic
-
-Q: Is Bob older than Tom?
-
-["When was Bob born?", "What is Tom's age?"]
-
-A: Yes, Bob is older than Tom. As Bob was born on 1984-01-01 and Tom is 30 years old.
-
-Q: What is their age difference?
-
-["What is Bob's age?", "What is Tom's age?"]
-
-A: Bob is {current_new_year.year - 1984 - 30} years older than Tom. As Bob is {current_new_year.year - 1984} years old and Tom is 30 years old.
-
-{chat_history}
-Q: {text}
-
-"""
 
     # Get Response from GPT
     response = completion_with_backoff(
@@ -191,31 +126,8 @@ def extract_search_type(text, model, api_key=None, temperature=0.5, max_tokens=1
     """
     Extract search type from user query using OpenAI's GPT
     """
-    # Initialize Variables
+    # Setup Prompt to extract search type
-    understand_primer = """
+    prompt = prompts.search_type + f"{text}\nA:"
-Objective: Extract search type from user query and return information as JSON
-
-Allowed search types are listed below:
-  - search-type=["notes","ledger","image","music"]
-
-Some examples are given below for reference:
-Q:What fiction book was I reading last week about AI starship?
-A:{ "search-type": "notes" }
-Q:Play some calm classical music?
-A:{ "search-type": "music" }
-Q:How much did I spend at Subway for dinner last time?
-A:{ "search-type": "ledger" }
-Q:What was that popular Sri lankan song that Alex had mentioned?
-A:{ "search-type": "music" }
-Q:Can you recommend a movie to watch from my notes?
-A:{ "search-type": "notes" }
-Q: When did I buy Groceries last?
-A:{ "search-type": "ledger" }
-Q:When did I go surfing last?
-A:{ "search-type": "notes" }"""
-
-    # Setup Prompt with Understand Primer
-    prompt = message_to_prompt(text, understand_primer, start_sequence="\nA:", restart_sequence="\nQ:")
     if verbose > 1:
         print(f"Message -> Prompt: {text} -> {prompt}")
 
@@ -241,36 +153,23 @@ def converse(references, user_query, conversation_log={}, model="gpt-3.5-turbo",
     Converse with user using OpenAI's ChatGPT
     """
     # Initialize Variables
+    current_date = datetime.now().strftime("%Y-%m-%d")
     compiled_references = "\n\n".join({f"# {item}" for item in references})
 
-    personality_primer = "You are Khoj, a friendly, smart and helpful personal assistant."
-    conversation_primers = {
-        "general": f"""
-Using your general knowledge and our past conversations as context, answer the following question.
-Current Date: {datetime.now().strftime("%Y-%m-%d")}
-
-Question: {user_query}
-""".strip(),
-        "notes": f"""
-Using the notes and our past conversations as context, answer the following question.
-Current Date: {datetime.now().strftime("%Y-%m-%d")}
-
-Notes:
-{compiled_references}
-
-Question: {user_query}
-""".strip(),
-    }
-
     # Get Conversation Primer appropriate to Conversation Type
     conversation_type = "general" if user_query.startswith("@general") or compiled_references.strip() == "" else "notes"
     logger.debug(f"Conversation Type: {conversation_type}")
-    conversation_primer = conversation_primers.get(conversation_type)
+    if conversation_type == "general":
+        conversation_primer = prompts.general_conversation.format(current_date=current_date, query=user_query)
+    else:
+        conversation_primer = prompts.notes_conversation.format(
+            current_date=current_date, query=user_query, references=compiled_references
+        )
 
     # Setup Prompt with Primer or Conversation History
     messages = generate_chatml_messages_with_context(
         conversation_primer,
-        personality_primer,
+        prompts.personality.format(),
         conversation_log,
         model,
     )

src/khoj/processor/conversation/prompts.py

@@ -0,0 +1,163 @@
+# External Packages
+from langchain.prompts import PromptTemplate
+
+
+## Personality
+## --
+personality = PromptTemplate.from_template("You are Khoj, a friendly, smart and helpful personal assistant.")
+
+
+## General Conversation
+## --
+general_conversation = PromptTemplate.from_template(
+    """
+Using your general knowledge and our past conversations as context, answer the following question.
+Current Date: {current_date}
+
+Question: {query}
+""".strip()
+)
+
+
+## Notes Conversation
+## --
+notes_conversation = PromptTemplate.from_template(
+    """
+Using the notes and our past conversations as context, answer the following question.
+Current Date: {current_date}
+
+Notes:
+{references}
+
+Question: {query}
+""".strip()
+)
+
+
+## Summarize Chat
+## --
+summarize_chat = PromptTemplate.from_template(
+    """
+You are an AI. Summarize the conversation below from your perspective:
+
+{text}
+
+Summarize the conversation from the AI's first-person perspective:"""
+)
+
+
+## Summarize Notes
+## --
+summarize_notes = PromptTemplate.from_template(
+    """
+Summarize the below notes about {user_query}:
+
+{text}
+
+Summarize the notes in second person perspective:"""
+)
+
+
+## Answer
+## --
+answer = PromptTemplate.from_template(
+    """
+You are a friendly, helpful personal assistant.
+Using the users notes below, answer their following question. If the answer is not contained within the notes, say "I don't know."
+
+Notes:
+{text}
+
+Question: {user_query}
+
+Answer (in second person):"""
+)
+
+
+## Extract Questions
+## --
+extract_questions = PromptTemplate.from_template(
+    """
+You are Khoj, an extremely smart and helpful search assistant with the ability to retrieve information from the users notes.
+- The user will provide their questions and answers to you for context.
+- Add as much context from the previous questions and answers as required into your search queries.
+- Break messages into multiple search queries when required to retrieve the relevant information.
+- Add date filters to your search queries from questions and answers when required to retrieve the relevant information.
+
+What searches, if any, will you need to perform to answer the users question?
+Provide search queries as a JSON list of strings
+Current Date: {current_date}
+
+Q: How was my trip to Cambodia?
+
+["How was my trip to Cambodia?"]
+
+A: The trip was amazing. I went to the Angkor Wat temple and it was beautiful.
+
+Q: Who did i visit that temple with?
+
+["Who did I visit the Angkor Wat Temple in Cambodia with?"]
+
+A: You visited the Angkor Wat Temple in Cambodia with Pablo, Namita and Xi.
+
+Q: What national parks did I go to last year?
+
+["National park I visited in {last_new_year} dt>=\\"{last_new_year_date}\\" dt<\\"{current_new_year_date}\\""]
+
+A: You visited the Grand Canyon and Yellowstone National Park in {last_new_year}.
+
+Q: How are you feeling today?
+
+[]
+
+A: I'm feeling a little bored. Helping you will hopefully make me feel better!
+
+Q: How many tennis balls fit in the back of a 2002 Honda Civic?
+
+["What is the size of a tennis ball?", "What is the trunk size of a 2002 Honda Civic?"]
+
+A: 1085 tennis balls will fit in the trunk of a Honda Civic
+
+Q: Is Bob older than Tom?
+
+["When was Bob born?", "What is Tom's age?"]
+
+A: Yes, Bob is older than Tom. As Bob was born on 1984-01-01 and Tom is 30 years old.
+
+Q: What is their age difference?
+
+["What is Bob's age?", "What is Tom's age?"]
+
+A: Bob is {bob_tom_age_difference} years older than Tom. As Bob is {bob_age} years old and Tom is 30 years old.
+
+{chat_history}
+Q: {text}
+
+"""
+)
+
+
+## Extract Search Type
+## --
+search_type = """
+Objective: Extract search type from user query and return information as JSON
+
+Allowed search types are listed below:
+  - search-type=["notes","ledger","image","music"]
+
+Some examples are given below for reference:
+Q:What fiction book was I reading last week about AI starship?
+A:{ "search-type": "notes" }
+Q:Play some calm classical music?
+A:{ "search-type": "music" }
+Q:How much did I spend at Subway for dinner last time?
+A:{ "search-type": "ledger" }
+Q:What was that popular Sri lankan song that Alex had mentioned?
+A:{ "search-type": "music" }
+Q:Can you recommend a movie to watch from my notes?
+A:{ "search-type": "notes" }
+Q:When did I buy Groceries last?
+A:{ "search-type": "ledger" }
+Q:When did I go surfing last?
+A:{ "search-type": "notes" }
+Q:"""