sij/khoj
1
0
Fork 0
mirror of https://github.com/khoj-ai/khoj.git synced 2024-11-27 17:35:07 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

Scale no. of chat history messages to use as context with max_prompt_size

Browse source 
Previously lookback turns was set to a static 2. But now that we
support more chat models, their prompt size vary considerably.

Make lookback_turns proportional to max_prompt_size. The truncate_messages
can remove messages if they exceed max_prompt_size later

This lets Khoj pass more of the chat history as context for models
with larger context window
This commit is contained in:
Debanjum Singh Solanky 2023-10-16 11:15:38 -07:00
parent 90e1d9e3d6
commit 644c3b787f
1 changed files with 13 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

24
src/khoj/processor/conversation/utils.py
View file

@ -3,6 +3,7 @@ import logging
from time import perf_counter
import json
from datetime import datetime
import queue
import tiktoken
# External packages
@ -10,7 +11,6 @@ from langchain.schema import ChatMessage
from transformers import AutoTokenizer
# Internal Packages
import queue
from khoj.utils.helpers import merge_dicts
@ -89,11 +89,22 @@ def generate_chatml_messages_with_context(
system_message,
conversation_log={},
model_name="gpt-3.5-turbo",
lookback_turns=2,
max_prompt_size=None,
tokenizer_name=None,
):
"""Generate messages for ChatGPT with context from previous conversation"""
# Set max prompt size from user config, pre-configured for model or to default prompt size
try:
max_prompt_size = max_prompt_size or model_to_prompt_size[model_name]
except:
max_prompt_size = 2000
logger.warning(
f"Fallback to default prompt size: {max_prompt_size}.\nConfigure max_prompt_size for unsupported model: {model_name} in Khoj settings to longer context window."
)
# Scale lookback turns proportional to max prompt size supported by model
lookback_turns = max_prompt_size // 750
# Extract Chat History for Context
chat_logs = []
for chat in conversation_log.get("chat", []):
@ -113,15 +124,6 @@ def generate_chatml_messages_with_context(
messages = user_chatml_message + rest_backnforths + system_chatml_message
# Set max prompt size from user config, pre-configured for model or to default prompt size
try:
max_prompt_size = max_prompt_size or model_to_prompt_size[model_name]
except:
max_prompt_size = 2000
logger.warning(
f"Fallback to default prompt size: {max_prompt_size}.\nConfigure max_prompt_size for unsupported model: {model_name} in Khoj settings to longer context window."
)
# Truncate oldest messages from conversation history until under max supported prompt size by model
messages = truncate_messages(messages, max_prompt_size, model_name, tokenizer_name)