Files
bDS2/lib/bds/ai/one_shot.ex

391 lines
12 KiB
Elixir

defmodule BDS.AI.OneShot do
@moduledoc false
alias BDS.AI.Chat
alias BDS.AI.OpenAICompatibleRuntime
alias BDS.AI.Runtime
alias BDS.Media.Media
alias BDS.MapUtils
alias BDS.Posts.Post
alias BDS.Repo
@default_max_output_tokens 16_384
@spec detect_language(String.t(), keyword()) :: {:ok, map()} | {:error, term()}
def detect_language(text, opts \\ []) when is_binary(text) and is_list(opts) do
run_one_shot(
:detect_language,
%{text: text},
opts,
fn json, usage ->
{:ok, %{language_code: json["language_code"], usage: usage}}
end
)
end
@spec analyze_taxonomy(map() | String.t(), keyword()) :: {:ok, map()} | {:error, term()}
def analyze_taxonomy(post_input, opts \\ []) when is_list(opts) do
with {:ok, post} <- normalize_post_input(post_input) do
run_one_shot(
:analyze_taxonomy,
post,
opts,
fn json, usage ->
{:ok,
%{
tags: json["tags"] || [],
categories: json["categories"] || [],
usage: usage
}}
end
)
end
end
@spec analyze_import_taxonomy(map(), map(), keyword()) :: {:ok, map()} | {:error, term()}
def analyze_import_taxonomy(import_terms, existing_terms, opts \\ [])
when is_map(import_terms) and is_map(existing_terms) and is_list(opts) do
payload = %{
import_categories: normalize_string_list(MapUtils.attr(import_terms, :categories)),
import_tags: normalize_string_list(MapUtils.attr(import_terms, :tags)),
existing_categories: normalize_string_list(MapUtils.attr(existing_terms, :categories)),
existing_tags: normalize_string_list(MapUtils.attr(existing_terms, :tags))
}
run_one_shot(
:import_taxonomy_mapping,
payload,
opts,
fn json, usage ->
{:ok,
%{
category_mappings:
filter_taxonomy_mapping_response(
json["categoryMappings"] || json["category_mappings"],
payload.import_categories,
payload.existing_categories
),
tag_mappings:
filter_taxonomy_mapping_response(
json["tagMappings"] || json["tag_mappings"],
payload.import_tags,
payload.existing_tags
),
usage: usage
}}
end
)
end
@spec analyze_post(map() | String.t(), keyword()) :: {:ok, map()} | {:error, term()}
def analyze_post(post_input, opts \\ []) when is_list(opts) do
with {:ok, post} <- normalize_post_input(post_input) do
run_one_shot(
:analyze_post,
post,
opts,
fn json, usage ->
{:ok,
%{
title: json["title"],
excerpt: json["excerpt"],
slug: json["slug"],
usage: usage
}}
end
)
end
end
@spec translate_post(map() | String.t(), String.t(), keyword()) ::
{:ok, map()} | {:error, term()}
def translate_post(post_input, target_language, opts \\ [])
when is_binary(target_language) and is_list(opts) do
with {:ok, post} <- normalize_post_input(post_input) do
run_one_shot(
:translate_post,
Map.put(post, :target_language, target_language),
opts,
fn json, usage ->
{:ok,
%{
title: json["title"],
excerpt: json["excerpt"],
content: json["content"],
usage: usage
}}
end
)
end
end
@spec analyze_image(map() | String.t(), keyword()) :: {:ok, map()} | {:error, term()}
def analyze_image(media_input, opts \\ []) when is_list(opts) do
with {:ok, media} <- normalize_media_input(media_input),
:ok <- ensure_image_media(media) do
run_one_shot(
:analyze_image,
media,
opts,
fn json, usage ->
{:ok,
%{
title: json["title"],
alt: json["alt"],
caption: json["caption"],
usage: usage
}}
end
)
end
end
@spec translate_media(map() | String.t(), String.t(), keyword()) ::
{:ok, map()} | {:error, term()}
def translate_media(media_input, target_language, opts \\ [])
when is_binary(target_language) and is_list(opts) do
with {:ok, media} <- normalize_media_input(media_input) do
run_one_shot(
:translate_media,
Map.put(media, :target_language, target_language),
opts,
fn json, usage ->
{:ok,
%{
title: json["title"],
alt: json["alt"],
caption: json["caption"],
usage: usage
}}
end
)
end
end
defp run_one_shot(operation, payload, opts, formatter) do
runtime = Keyword.get(opts, :runtime, OpenAICompatibleRuntime)
with {:ok, endpoint, model, mode} <- Runtime.resolve_target(operation, opts),
:ok <- Runtime.validate_target(operation, model, mode),
request <- build_one_shot_request(operation, payload, model),
{:ok, response} <-
runtime.generate(Runtime.endpoint_with_model(endpoint, model), request, opts),
{:ok, json} <- extract_json_response(response),
usage <- Chat.normalize_usage(response.usage),
{:ok, result} <- formatter.(json, usage) do
{:ok, result}
end
end
defp build_one_shot_request(operation, payload, model) do
%{
operation: operation,
model: model,
max_output_tokens: @default_max_output_tokens,
messages: [
%{"role" => "system", "content" => one_shot_system_prompt(operation)},
%{"role" => "user", "content" => one_shot_user_content(operation, payload)}
]
}
end
defp one_shot_system_prompt(:detect_language) do
"Return JSON with exactly one key: language_code."
end
defp one_shot_system_prompt(:analyze_taxonomy) do
"Return JSON with keys tags and categories, each an array of short strings."
end
defp one_shot_system_prompt(:import_taxonomy_mapping) do
"You are helping import WordPress taxonomy into an existing blog. Return JSON with exactly two keys: categoryMappings and tagMappings. Each value must be an object mapping imported term names to existing project term names. Only map when the imported term should reuse an existing term to avoid duplicates. Do not invent target terms. Leave unmapped items out of the objects."
end
defp one_shot_system_prompt(:analyze_post) do
"Return JSON with keys title, excerpt, and slug."
end
defp one_shot_system_prompt(:translate_post) do
"Return JSON with keys title, excerpt, and content. Preserve Markdown structure."
end
defp one_shot_system_prompt(:analyze_image) do
"Return JSON with keys title, alt, and caption for the provided image."
end
defp one_shot_system_prompt(:translate_media) do
"Return JSON with keys title, alt, and caption translated to the requested language."
end
defp one_shot_user_content(:detect_language, %{text: text}) do
"Detect the language of this text: #{text}"
end
defp one_shot_user_content(:analyze_taxonomy, post) do
"Suggest categories and tags for the following post.\nTitle: #{post.title}\nExcerpt: #{post.excerpt}\nContent: #{truncate_text(post.content, 2000)}"
end
defp one_shot_user_content(:import_taxonomy_mapping, payload) do
[
"Analyze these imported taxonomy terms and suggest which ones should map to existing project terms.",
"",
"Imported categories:",
Enum.join(payload.import_categories, ", "),
"",
"Imported tags:",
Enum.join(payload.import_tags, ", "),
"",
"Existing project categories:",
Enum.join(payload.existing_categories, ", "),
"",
"Existing project tags:",
Enum.join(payload.existing_tags, ", "),
"",
"Return JSON only."
]
|> Enum.join("\n")
end
defp one_shot_user_content(:analyze_post, post) do
"Suggest an improved title, excerpt, and slug.\nTitle: #{post.title}\nExcerpt: #{post.excerpt}\nContent: #{truncate_text(post.content, 2000)}"
end
defp one_shot_user_content(:translate_post, post) do
"Translate this post to #{post.target_language}.\nTitle: #{post.title}\nExcerpt: #{post.excerpt}\nContent: #{post.content}"
end
defp one_shot_user_content(:analyze_image, media) do
[
%{
"type" => "text",
"text" => "Analyze this image and return title, alt text, and caption."
},
%{"type" => "image_url", "image_url" => %{"url" => media.image_url}}
]
end
defp one_shot_user_content(:translate_media, media) do
"Translate this media metadata to #{media.target_language}.\nTitle: #{media.title}\nAlt: #{media.alt}\nCaption: #{media.caption}"
end
defp extract_json_response(%{json: json}) when is_map(json), do: {:ok, json}
defp extract_json_response(%{content: content}) when is_binary(content) do
case Jason.decode(content) do
{:ok, json} when is_map(json) -> {:ok, json}
_other -> {:error, %{kind: :invalid_json_response}}
end
end
defp extract_json_response(_response), do: {:error, %{kind: :invalid_json_response}}
defp normalize_post_input(%Post{} = post) do
{:ok, %{title: post.title || "", excerpt: post.excerpt || "", content: post.content || ""}}
end
defp normalize_post_input(post_id) when is_binary(post_id) do
case Repo.get(Post, post_id) do
nil -> {:error, :not_found}
post -> normalize_post_input(post)
end
end
defp normalize_post_input(attrs) when is_map(attrs) do
{:ok,
%{
title: MapUtils.attr(attrs, :title) || "",
excerpt: MapUtils.attr(attrs, :excerpt) || "",
content: MapUtils.attr(attrs, :content) || ""
}}
end
defp normalize_media_input(%Media{} = media) do
{:ok,
%{
mime_type: media.mime_type,
title: media.title || "",
alt: media.alt || "",
caption: media.caption || "",
image_url: Map.get(media, :image_url) || media_path_to_file_url(media.file_path)
}}
end
defp normalize_media_input(media_id) when is_binary(media_id) do
case Repo.get(Media, media_id) do
nil -> {:error, :not_found}
media -> normalize_media_input(media)
end
end
defp normalize_media_input(attrs) when is_map(attrs) do
{:ok,
%{
mime_type: MapUtils.attr(attrs, :mime_type),
title: MapUtils.attr(attrs, :title) || "",
alt: MapUtils.attr(attrs, :alt) || "",
caption: MapUtils.attr(attrs, :caption) || "",
image_url: MapUtils.attr(attrs, :image_url)
}}
end
defp ensure_image_media(%{mime_type: "image/" <> _rest}), do: :ok
defp ensure_image_media(_media), do: {:error, %{kind: :invalid_media_type}}
defp media_path_to_file_url(nil), do: nil
defp media_path_to_file_url(path), do: "file://" <> path
defp normalize_string_list(values) do
values
|> List.wrap()
|> Enum.map(&to_string/1)
|> Enum.map(&String.trim/1)
|> Enum.reject(&(&1 == ""))
|> Enum.uniq()
end
defp filter_taxonomy_mapping_response(mappings, import_terms, existing_terms)
when is_map(mappings) do
import_lookup = canonical_term_lookup(import_terms)
existing_lookup = canonical_term_lookup(existing_terms)
Enum.reduce(mappings, %{}, fn {source, target}, acc ->
with {:ok, canonical_source} <- resolve_canonical_term(source, import_lookup),
{:ok, canonical_target} <- resolve_canonical_term(target, existing_lookup) do
Map.put(acc, canonical_source, canonical_target)
else
_other -> acc
end
end)
end
defp filter_taxonomy_mapping_response(_mappings, _import_terms, _existing_terms), do: %{}
defp canonical_term_lookup(terms) do
Map.new(terms, fn term -> {normalize_term(term), term} end)
end
defp resolve_canonical_term(term, lookup) do
case Map.get(lookup, normalize_term(term)) do
nil -> :error
canonical -> {:ok, canonical}
end
end
defp normalize_term(term) do
term
|> to_string()
|> String.trim()
|> String.downcase()
end
defp truncate_text(nil, _max_length), do: ""
defp truncate_text(text, max_length) when is_binary(text) do
if String.length(text) <= max_length do
text
else
String.slice(text, 0, max_length)
end
end
end