mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-08-07 13:43:57 +00:00
e817b7dc11
This re-implements tool support in DiscourseAi::Completions::Llm #generate
Previously tool support was always returned via XML and it would be the responsibility of the caller to parse XML
New implementation has the endpoints return ToolCall objects.
Additionally this simplifies the Llm endpoint interface and gives it more clarity. Llms must implement
decode, decode_chunk (for streaming)
It is the implementers responsibility to figure out how to decode chunks, base no longer implements. To make this easy we ship a flexible json decoder which is easy to wire up.
Also (new)
Better debugging for PMs, we now have a next / previous button to see all the Llm messages associated with a PM
Token accounting is fixed for vllm (we were not correctly counting tokens)
125 lines
3.5 KiB
Ruby
125 lines
3.5 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
# This class can be used to process a stream of text that may contain XML tool
|
|
# calls.
|
|
# It will return either text or ToolCall objects.
|
|
|
|
module DiscourseAi
|
|
module Completions
|
|
class XmlToolProcessor
|
|
def initialize
|
|
@buffer = +""
|
|
@function_buffer = +""
|
|
@should_cancel = false
|
|
@in_tool = false
|
|
end
|
|
|
|
def <<(text)
|
|
@buffer << text
|
|
result = []
|
|
|
|
if !@in_tool
|
|
# double check if we are clearly in a tool
|
|
search_length = text.length + 20
|
|
search_string = @buffer[-search_length..-1] || @buffer
|
|
|
|
index = search_string.rindex("<function_calls>")
|
|
@in_tool = !!index
|
|
if @in_tool
|
|
@function_buffer = @buffer[index..-1]
|
|
text_index = text.rindex("<function_calls>")
|
|
result << text[0..text_index - 1].strip if text_index && text_index > 0
|
|
end
|
|
else
|
|
@function_buffer << text
|
|
end
|
|
|
|
if !@in_tool
|
|
if maybe_has_tool?(@buffer)
|
|
split_index = text.rindex("<").to_i - 1
|
|
if split_index >= 0
|
|
@function_buffer = text[split_index + 1..-1] || ""
|
|
text = text[0..split_index] || ""
|
|
else
|
|
@function_buffer << text
|
|
text = ""
|
|
end
|
|
else
|
|
if @function_buffer.length > 0
|
|
result << @function_buffer
|
|
@function_buffer = +""
|
|
end
|
|
end
|
|
|
|
result << text if text.length > 0
|
|
else
|
|
@should_cancel = true if text.include?("</function_calls>")
|
|
end
|
|
|
|
result
|
|
end
|
|
|
|
def finish
|
|
return [] if @function_buffer.blank?
|
|
|
|
xml = Nokogiri::HTML5.fragment(@function_buffer)
|
|
normalize_function_ids!(xml)
|
|
last_invoke = xml.at("invoke:last")
|
|
if last_invoke
|
|
last_invoke.next_sibling.remove while last_invoke.next_sibling
|
|
xml.at("invoke:last").add_next_sibling("\n") if !last_invoke.next_sibling
|
|
end
|
|
|
|
xml
|
|
.css("invoke")
|
|
.map do |invoke|
|
|
tool_name = invoke.at("tool_name").content.force_encoding("UTF-8")
|
|
tool_id = invoke.at("tool_id").content.force_encoding("UTF-8")
|
|
parameters = {}
|
|
invoke
|
|
.at("parameters")
|
|
&.children
|
|
&.each do |node|
|
|
next if node.text?
|
|
name = node.name
|
|
value = node.content.to_s
|
|
parameters[name.to_sym] = value.to_s.force_encoding("UTF-8")
|
|
end
|
|
ToolCall.new(id: tool_id, name: tool_name, parameters: parameters)
|
|
end
|
|
end
|
|
|
|
def should_cancel?
|
|
@should_cancel
|
|
end
|
|
|
|
private
|
|
|
|
def normalize_function_ids!(function_buffer)
|
|
function_buffer
|
|
.css("invoke")
|
|
.each_with_index do |invoke, index|
|
|
if invoke.at("tool_id")
|
|
invoke.at("tool_id").content = "tool_#{index}" if invoke.at("tool_id").content.blank?
|
|
else
|
|
invoke.add_child("<tool_id>tool_#{index}</tool_id>\n") if !invoke.at("tool_id")
|
|
end
|
|
end
|
|
end
|
|
|
|
def maybe_has_tool?(text)
|
|
# 16 is the length of function calls
|
|
substring = text[-16..-1] || text
|
|
split = substring.split("<")
|
|
|
|
if split.length > 1
|
|
match = "<" + split.last
|
|
"<function_calls>".start_with?(match)
|
|
else
|
|
substring.ends_with?("<")
|
|
end
|
|
end
|
|
end
|
|
end
|
|
end
|