workflows

name	workflows
description	The orchestration pipelines and Jupyter notebook workflows for generating language learning materials.

Audio Language Trainer Workflows

Overview

The system relies on Jupyter notebooks to orchestrate the generation and packaging of language learning materials. The standard pipeline is: Vocab/Verbs -> English Phrases -> Translated/Media Phrases -> Anki Decks / Stories.

1. Phrase Generation (01 Flashcards - create phrases.ipynb)

• Input: Text files containing lists of target verbs and vocabulary (e.g., LM1000_verbs.txt).
• Process:
  1. An LLM (src.phrases.generation) generates short, simple English phrases combining these verbs and vocab.
  2. For each generated string, a Phrase object is created.
  3. The phrase is translated into the target language (via Google Translate + Anthropic Claude refinement).
  4. An image is generated (via Imagen, Stability, or DeepAI).
  5. Text-to-Speech (TTS) audio is generated for both 'flashcard' and 'story' contexts.
  6. The Phrase and its multimedia are uploaded to Firestore and GCS.

2. Anki Deck Creation (04 Anki Deck.ipynb)

• Input: A collection of generated phrases in Firestore/GCS.
• Process:
  1. Queries Firestore for phrases belonging to a specific collection/deck (e.g., LM1000 Pack01).
  2. Uses src.anki_tools.create_anki_deck to bundle the phrases, target language audio, images, and translations into an Anki package.
  3. Output: .apkg files saved locally for import into Anki.

3. Story Generation (05 - generate a story from vocab.ipynb)

• Input: A list of learned phrases from a deck.
• Process:
  1. Extracts verbs/vocab from the phrases.
  2. Calls an LLM (src.llm_tools.story_generation) to generate a dialogue-based story incorporating these phrases.
  3. Instantiates a Story object and saves the structure.
  4. Publishes the story (story.publish_story), which generates combined audio files, story-specific translations, and an index.html file on GCS for web access.