mirror of
https://gh.wpcy.net/https://github.com/discourse/discourse.git
synced 2026-05-27 20:52:53 +08:00
b8abe100c5
Introduce an "agentic" execution mode as an alternative to the default fixed-turn/tool-limit approach. In agentic mode, personas use a configurable token budget (`max_turn_tokens`) to govern how long a tool-use session can run, with automatic context compression when the conversation exceeds a configurable threshold percentage (`compression_threshold`) of the model's context window. Key changes: - Add `execution_mode`, `max_turn_tokens`, and `compression_threshold` columns to `ai_personas` via migration - Refactor `Bot#reply` to support token-budget loop control with a thread-local token accumulator, budget exhaustion hints, and a safety valve at 100 completions - Add `maybe_compress_context` which summarizes middle conversation messages when token usage crosses the compression threshold, preserving system prompt and recent tail messages - Update `StreamReplyCustomToolsSession` to track accumulated tokens across rounds and handle budget exhaustion in the custom tools path - Discount cached tokens (Anthropic) in the token accumulator to avoid over-counting reused KV cache prefixes - Update persona editor UI with execution mode selector and conditional fields (agentic shows token budget/compression; default shows max context posts)
124 lines
3.1 KiB
Ruby
Vendored
124 lines
3.1 KiB
Ruby
Vendored
# frozen_string_literal: true
|
|
|
|
class DiscourseAi::Utils::PdfToText
|
|
MAX_PDF_SIZE = 100.megabytes
|
|
|
|
class Reader
|
|
def initialize(upload:, user: nil, llm_model: nil, execution_context: nil)
|
|
@extractor =
|
|
DiscourseAi::Utils::PdfToText.new(
|
|
upload: upload,
|
|
user: user,
|
|
llm_model: llm_model,
|
|
execution_context:,
|
|
)
|
|
@enumerator = create_enumerator
|
|
@buffer = +""
|
|
end
|
|
|
|
def read(length)
|
|
return @buffer.slice!(0, length) if !@buffer.empty?
|
|
|
|
begin
|
|
@buffer << @enumerator.next
|
|
rescue StopIteration
|
|
return nil
|
|
end
|
|
|
|
@buffer.slice!(0, length)
|
|
end
|
|
|
|
private
|
|
|
|
def create_enumerator
|
|
Enumerator.new { |yielder| @extractor.extract_text { |chunk| yielder.yield(chunk || "") } }
|
|
end
|
|
end
|
|
|
|
attr_reader :upload
|
|
|
|
def self.as_fake_file(upload:, user: nil, llm_model: nil, execution_context: nil)
|
|
Reader.new(upload: upload, user: user, llm_model: llm_model, execution_context:)
|
|
end
|
|
|
|
def initialize(upload:, user: nil, llm_model: nil, execution_context: nil)
|
|
@upload = upload
|
|
@user = user
|
|
@llm_model = llm_model
|
|
@execution_context = execution_context
|
|
end
|
|
|
|
def extract_text
|
|
pdf_path =
|
|
if upload.local?
|
|
Discourse.store.path_for(upload)
|
|
else
|
|
Discourse.store.download(upload, max_file_size_kb: MAX_PDF_SIZE)
|
|
end
|
|
|
|
raise Discourse::InvalidParameters.new("Failed to download PDF") if pdf_path.nil?
|
|
|
|
require "pdf/reader"
|
|
|
|
page_number = 0
|
|
PDF::Reader.open(pdf_path) do |reader|
|
|
reader.pages.each do |page|
|
|
page_number += 1
|
|
llm_decorate(page_number: page_number, text: page.text, pdf_path: pdf_path) do |chunk|
|
|
yield chunk
|
|
end
|
|
end
|
|
end
|
|
end
|
|
|
|
def llm_decorate(page_number:, text:, pdf_path:)
|
|
raise "Must be called with block" if !block_given?
|
|
if !@llm_model
|
|
yield text
|
|
return
|
|
end
|
|
|
|
begin
|
|
temp_dir = Dir.mktmpdir("discourse-pdf-#{SecureRandom.hex(8)}")
|
|
output_path = File.join(temp_dir, "page-#{page_number}.png")
|
|
|
|
# Extract specific page using ImageMagick
|
|
# image magick uses 0 based page numbers
|
|
command = [
|
|
"magick",
|
|
"-density",
|
|
"300",
|
|
"#{pdf_path}[#{page_number - 1}]",
|
|
"-background",
|
|
"white",
|
|
"-auto-orient",
|
|
"-quality",
|
|
"85",
|
|
output_path,
|
|
]
|
|
|
|
Discourse::Utils.execute_command(
|
|
*command,
|
|
failure_message: "Failed to convert PDF page #{page_number} to image",
|
|
timeout: 30,
|
|
)
|
|
|
|
# TODO - we are creating leftover uploads, they will be cleaned up
|
|
# but maybe we should just keep them around?
|
|
upload =
|
|
UploadCreator.new(File.open(output_path), "page-#{page_number}.png").create_for(@user&.id)
|
|
|
|
DiscourseAi::Utils::ImageToText
|
|
.new(
|
|
upload: upload,
|
|
llm_model: @llm_model,
|
|
user: @user,
|
|
guidance_text: text,
|
|
execution_context: @execution_context,
|
|
)
|
|
.extract_text { |chunk| yield chunk }
|
|
ensure
|
|
FileUtils.rm_rf(temp_dir) if temp_dir
|
|
end
|
|
end
|
|
end
|
