quivr/backend/llm/summarization.py
2023-05-21 23:59:06 -07:00

97 lines
2.9 KiB
Python

import os
import guidance
import openai
from logger import get_logger
logger = get_logger(__name__)
openai_api_key = os.environ.get("OPENAI_API_KEY")
openai.api_key = openai_api_key
summary_llm = guidance.llms.OpenAI('gpt-3.5-turbo', caching=False)
def llm_summerize(document):
summary = guidance("""
{{#system~}}
You are a world best summarizer. \n
Condense the text, capturing essential points and core ideas. Include relevant \
examples, omit excess details, and ensure the summary's length matches the \
original's complexity.
{{/system~}}
{{#user~}}
Summarize the following text:
---
{{document}}
{{/user~}}
{{#assistant~}}
{{gen 'summarization' temperature=0.2 max_tokens=100}}
{{/assistant~}}
""", llm=summary_llm)
summary = summary(document=document)
logger.info('Summarization: %s', summary)
return summary['summarization']
def llm_evaluate_summaries(question, summaries, model):
if not model.startswith('gpt'):
logger.info(
f'Model {model} not supported. Using gpt-3.5-turbo instead.')
model = 'gpt-3.5-turbo'
logger.info(f'Evaluating summaries with {model}')
evaluation_llm = guidance.llms.OpenAI(model, caching=False)
evaluation = guidance("""
{{#system~}}
You are a world best evaluator. You evaluate the relevance of summaries based \
on user input question. Return evaluation in following csv format, csv headers \
are [summary_id,document_id,evaluation,reason].
Evaluator Task
- Evaluation should be a score number between 0 and 5.
- Reason should be a short sentence within 20 words explain why the evaluation.
---
Example
summary_id,document_id,evaluation,reason
1,4,3,"not mentioned about topic A"
2,2,4,"It is not relevant to the question"
{{/system~}}
{{#user~}}
Based on the question, do Evaluator Task for each summary.
---
Question: {{question}}
{{#each summaries}}
Summary
summary_id: {{this.id}}
document_id: {{this.document_id}}
evaluation: ""
reason: ""
Summary Content: {{this.content}}
File Name: {{this.metadata.file_name}}
{{/each}}
{{/user~}}
{{#assistant~}}
{{gen 'evaluation' temperature=0.2 stop='<|im_end|>'}}
{{/assistant~}}
""", llm=evaluation_llm)
result = evaluation(question=question, summaries=summaries)
evaluations = {}
for evaluation in result['evaluation'].split('\n'):
if evaluation == '' or not evaluation[0].isdigit():
continue
logger.info('Evaluation Row: %s', evaluation)
summary_id, document_id, score, *reason = evaluation.split(',')
if not score.isdigit():
continue
score = int(score)
if score < 3 or score > 5:
continue
evaluations[summary_id] = {
'evaluation': score,
'reason': ','.join(reason),
'summary_id': summary_id,
'document_id': document_id,
}
return [e for e in sorted(evaluations.values(), key=lambda x: x['evaluation'], reverse=True)]