LLM chains¶

The lang_tools.llm package wraps llm-core's StructuredLLMChain for the seven recurring tasks of the language-learning stack. Every chain ships with a versioned Jinja prompt under src/lang_tools/prompts/<name>/v1.jinja and a build_*_chain(chat_config, ...) factory.

Each factory accepts:

chat_config: any llm-core ChatConfig (e.g. ChatOpenAIConfig(api_key=...)).
base_prompt_fol: optional prompts root override.
version: prompt version (default "auto", picks the highest vN).

from llm_core.chat.config.openai import ChatOpenAIConfig
from lang_tools.llm import build_translation_chain, TranslationInput

chain = build_translation_chain(ChatOpenAIConfig())
result = chain.invoke(TranslationInput(text="Bom dia",
                                       source_language="pt",
                                       target_language="en"))
print(result.translated_text)

Available chains¶

Factory	Input	Output	Purpose
`build_translation_chain`	`TranslationInput`	`TranslationOutput`	Translate a single string
`build_conversation_chain`	`ConversationInput`	`ConversationOutput`	Generate a multi-turn bilingual dialogue
`build_tutor_chain`	`TutorInput`	`TutorOutput`	Tutor reply with `correction` + `conversation` blocks
`build_topic_suggestion_chain`	`TopicSuggestionInput`	`TopicSuggestionOutput`	Suggest topics for practice
`build_paragraph_splitter_chain`	`SplitterInput`	`SplitterOutput`	Split text into reconstruction-friendly portions
`build_greeting_chain`	`GreetingInput`	`GreetingOutput`	Open a tutor conversation
`build_word_generator_chain`	`WordGeneratorInput`	`WordGeneratorOutput`	Generate themed vocabulary on demand

StructuredLLMChain validates that every Jinja variable in the prompt matches a field of the input model at construction time, so a prompt edit that drops or renames a variable surfaces immediately.