mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-07-24 15:03:26 +00:00
eb7fff3a55
* FEATURE: Add support for StableBeluga and Upstage Llama2 instruct This means we support all models in the top3 of the Open LLM Leaderboard Since some of those models have RoPE, we now have a setting so you can customize the token limit depending which model you use.
107 lines
3.0 KiB
Ruby
107 lines
3.0 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Summarization
|
|
module Models
|
|
class Llama2 < Base
|
|
def display_name
|
|
"Llama2's #{SiteSetting.ai_hugging_face_model_display_name.presence || model}"
|
|
end
|
|
|
|
def correctly_configured?
|
|
SiteSetting.ai_hugging_face_api_url.present?
|
|
end
|
|
|
|
def configuration_hint
|
|
I18n.t(
|
|
"discourse_ai.summarization.configuration_hint",
|
|
count: 1,
|
|
setting: "ai_hugging_face_api_url",
|
|
)
|
|
end
|
|
|
|
def concatenate_summaries(summaries)
|
|
completion(<<~TEXT)
|
|
[INST] <<SYS>>
|
|
You are a helpful bot
|
|
<</SYS>>
|
|
|
|
Concatenate these disjoint summaries, creating a cohesive narrative:
|
|
#{summaries.join("\n")} [/INST]
|
|
TEXT
|
|
end
|
|
|
|
def summarize_with_truncation(contents, opts)
|
|
text_to_summarize = contents.map { |c| format_content_item(c) }.join
|
|
truncated_content = tokenizer.truncate(text_to_summarize, available_tokens)
|
|
|
|
completion(<<~TEXT)
|
|
[INST] <<SYS>>
|
|
#{build_base_prompt(opts)}
|
|
<</SYS>>
|
|
|
|
Summarize the following in up to 400 words:
|
|
#{truncated_content} [/INST]
|
|
Here is a summary of the above topic:
|
|
TEXT
|
|
end
|
|
|
|
def summarize_single(chunk_text, opts)
|
|
summarize_chunk(chunk_text, opts.merge(single_chunk: true))
|
|
end
|
|
|
|
private
|
|
|
|
def summarize_chunk(chunk_text, opts)
|
|
summary_instruction =
|
|
if opts[:single_chunk]
|
|
"Summarize the following forum discussion, creating a cohesive narrative:"
|
|
else
|
|
"Summarize the following in up to 400 words:"
|
|
end
|
|
|
|
completion(<<~TEXT)
|
|
[INST] <<SYS>>
|
|
#{build_base_prompt(opts)}
|
|
<</SYS>>
|
|
|
|
#{summary_instruction}
|
|
#{chunk_text} [/INST]
|
|
Here is a summary of the above topic:
|
|
TEXT
|
|
end
|
|
|
|
def build_base_prompt(opts)
|
|
base_prompt = <<~TEXT
|
|
You are a summarization bot.
|
|
You effectively summarise any text and reply ONLY with ONLY the summarized text.
|
|
You condense it into a shorter version.
|
|
You understand and generate Discourse forum Markdown.
|
|
TEXT
|
|
|
|
if opts[:resource_path]
|
|
base_prompt +=
|
|
"Try generating links as well the format is #{opts[:resource_path]}. eg: [ref](#{opts[:resource_path]}/77)\n"
|
|
end
|
|
|
|
base_prompt += "The discussion title is: #{opts[:content_title]}.\n" if opts[
|
|
:content_title
|
|
]
|
|
|
|
base_prompt
|
|
end
|
|
|
|
def completion(prompt)
|
|
::DiscourseAi::Inference::HuggingFaceTextGeneration.perform!(prompt, model).dig(
|
|
:generated_text,
|
|
)
|
|
end
|
|
|
|
def tokenizer
|
|
DiscourseAi::Tokenizer::Llama2Tokenizer
|
|
end
|
|
end
|
|
end
|
|
end
|
|
end
|