Sam e817b7dc11
FEATURE: improve tool support (#904)
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)
2024-11-12 08:14:30 +11:00

232 lines
6.4 KiB
Ruby

# frozen_string_literal: true
module DiscourseAi
module Completions
module Endpoints
class Gemini < Base
def self.can_contact?(model_provider)
model_provider == "google"
end
def default_options
# the default setting is a problem, it blocks too much
categories = %w[HARASSMENT SEXUALLY_EXPLICIT HATE_SPEECH DANGEROUS_CONTENT]
safety_settings =
categories.map do |category|
{ category: "HARM_CATEGORY_#{category}", threshold: "BLOCK_NONE" }
end
{ generationConfig: {}, safetySettings: safety_settings }
end
def normalize_model_params(model_params)
model_params = model_params.dup
if model_params[:stop_sequences]
model_params[:stopSequences] = model_params.delete(:stop_sequences)
end
if model_params[:max_tokens]
model_params[:maxOutputTokens] = model_params.delete(:max_tokens)
end
model_params[:topP] = model_params.delete(:top_p) if model_params[:top_p]
# temperature already supported
model_params
end
def provider_id
AiApiAuditLog::Provider::Gemini
end
private
def model_uri
url = llm_model.url
key = llm_model.api_key
if @streaming_mode
url = "#{url}:streamGenerateContent?key=#{key}&alt=sse"
else
url = "#{url}:generateContent?key=#{key}"
end
URI(url)
end
def prepare_payload(prompt, model_params, dialect)
tools = dialect.tools
payload = default_options.merge(contents: prompt[:messages])
payload[:systemInstruction] = {
role: "system",
parts: [{ text: prompt[:system_instruction].to_s }],
} if prompt[:system_instruction].present?
if tools.present?
payload[:tools] = tools
function_calling_config = { mode: "AUTO" }
if dialect.tool_choice.present?
function_calling_config = {
mode: "ANY",
allowed_function_names: [dialect.tool_choice],
}
end
payload[:tool_config] = { function_calling_config: function_calling_config }
end
payload[:generationConfig].merge!(model_params) if model_params.present?
payload
end
def prepare_request(payload)
headers = { "Content-Type" => "application/json" }
Net::HTTP::Post.new(model_uri, headers).tap { |r| r.body = payload }
end
def extract_completion_from(response_raw)
parsed =
if @streaming_mode
response_raw
else
JSON.parse(response_raw, symbolize_names: true)
end
response_h = parsed.dig(:candidates, 0, :content, :parts, 0)
if response_h
@has_function_call ||= response_h.dig(:functionCall).present?
@has_function_call ? response_h.dig(:functionCall) : response_h.dig(:text)
end
end
class GeminiStreamingDecoder
def initialize
@buffer = +""
end
def decode(str)
@buffer << str
lines = @buffer.split(/\r?\n\r?\n/)
keep_last = false
decoded =
lines
.map do |line|
if line.start_with?("data: {")
begin
JSON.parse(line[6..-1], symbolize_names: true)
rescue JSON::ParserError
keep_last = line
nil
end
else
keep_last = line
nil
end
end
.compact
if keep_last
@buffer = +(keep_last)
else
@buffer = +""
end
decoded
end
end
def decode(chunk)
json = JSON.parse(chunk, symbolize_names: true)
idx = -1
json
.dig(:candidates, 0, :content, :parts)
.map do |part|
if part[:functionCall]
idx += 1
ToolCall.new(
id: "tool_#{idx}",
name: part[:functionCall][:name],
parameters: part[:functionCall][:args],
)
else
part = part[:text]
if part != ""
part
else
nil
end
end
end
end
def decode_chunk(chunk)
@tool_index ||= -1
streaming_decoder
.decode(chunk)
.map do |parsed|
update_usage(parsed)
parsed
.dig(:candidates, 0, :content, :parts)
.map do |part|
if part[:text]
part = part[:text]
if part != ""
part
else
nil
end
elsif part[:functionCall]
@tool_index += 1
ToolCall.new(
id: "tool_#{@tool_index}",
name: part[:functionCall][:name],
parameters: part[:functionCall][:args],
)
end
end
end
.flatten
.compact
end
def update_usage(parsed)
usage = parsed.dig(:usageMetadata)
if usage
if prompt_token_count = usage[:promptTokenCount]
@prompt_token_count = prompt_token_count
end
if candidate_token_count = usage[:candidatesTokenCount]
@candidate_token_count = candidate_token_count
end
end
end
def final_log_update(log)
log.request_tokens = @prompt_token_count if @prompt_token_count
log.response_tokens = @candidate_token_count if @candidate_token_count
end
def streaming_decoder
@decoder ||= GeminiStreamingDecoder.new
end
def extract_prompt_for_tokenizer(prompt)
prompt.to_s
end
def xml_tools_enabled?
false
end
end
end
end
end