klbr/khoj
1
0
Fork 0
mirror of https://github.com/khoaliber/khoj.git synced 2026-03-06 05:39:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Switch default tokenizer to tiktoken as more widely used

Browse Source 
The tiktoken BPE based tokenizers seem more widely used these days.

Fallback to gpt-4o tiktoken tokenizer to count tokens for context
stuffing
This commit is contained in:
Debanjum Singh Solanky
2024-10-06 15:45:13 -07:00
parent 2b8f7f3efb
commit 516472a8d5
1 changed files with 3 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/khoj/processor/conversation/utils.py
View File

@@ -223,7 +223,7 @@ def truncate_messages(
) -> list[ChatMessage]:
"""Truncate messages to fit within max prompt size supported by model"""
default_tokenizer = "hf-internal-testing/llama-tokenizer"
default_tokenizer = "gpt-4o"
try:
if loaded_model:
@@ -240,13 +240,9 @@ def truncate_messages(
else:
encoder = download_model(model_name).tokenizer()
except:
if default_tokenizer in state.pretrained_tokenizers:
encoder = state.pretrained_tokenizers[default_tokenizer]
else:
encoder = AutoTokenizer.from_pretrained(default_tokenizer)
state.pretrained_tokenizers[default_tokenizer] = encoder
encoder = tiktoken.encoding_for_model(default_tokenizer)
logger.debug(
f"Fallback to default chat model tokenizer: {tokenizer_name}.\nConfigure tokenizer for unsupported model: {model_name} in Khoj settings to improve context stuffing."
f"Fallback to default chat model tokenizer: {default_tokenizer}.\nConfigure tokenizer for model: {model_name} in Khoj settings to improve context stuffing."
)
# Extract system message from messages