AI & Chat
Shiori can use a large language model for two things: sentence explanations in the reader and conversation practice in the Production view. The LLM is optional — every other feature works without one.
Providers
Configure a backend under Settings → AI. Three providers are supported:
| Provider | What you need | Notes |
|---|---|---|
| Anthropic | API key + model name | Key is stored locally in settings.json and sent only to the Anthropic API. Leave the field empty to use the ANTHROPIC_API_KEY environment variable instead. |
| Ollama (local) | Ollama installed and running | Models run entirely on your machine; nothing leaves it. |
| Custom endpoint | Base URL, model, optional API key | Any server speaking the OpenAI chat-completions dialect: LM Studio, llama.cpp server, vLLM, or a cloud provider. |
The bottom of the page shows the currently active backend, or "none" if the configuration is incomplete.
Ollama details
- Detection — the settings page probes the server (default
http://localhost:11434) and shows its status: running with version and installed-model count, or "not reachable" with a pointer to install Ollama from ollama.com. A refresh button re-probes. - Model picker — when the server is reachable, a dropdown lists every installed model with its size on disk and parameter count.
- In-app pulls — type a model name (e.g.
qwen3:8b) and press Pull; download progress streams into a progress bar in the settings page. Japanese-capable suggestions: qwen3, gemma3, llama3.1-swallow. - Fully offline — once a model is pulled, explanations and chat work with no network at all.
- The first request after a cold start loads the model into memory and can take tens of seconds; this is normal.
The server URL field accepts a remote address too (e.g. an Ollama instance on another machine on your network).
What the LLM powers
- Sentence explanations in the Reading view: an explanation of the current sentence's grammar and structure, on demand.
- Conversation practice in the Production view, described below.
If no backend is configured, the Production input area is replaced by a note pointing to Settings → AI.
Production chat
Production is a chat with a native-speaker persona. The design separates conversation from correction:
- The partner converses, never corrects. It replies only in Japanese, reacts, asks follow-ups, and responds to what you meant — corrections never appear inside its replies.
- Corrections come back as a paper-style write-up of your own messages. Each model call returns both the reply and a set of annotations on your latest message, rendered as colored underlines:
| Underline | Meaning |
|---|---|
| Red | Grammatically wrong |
| Orange | Correct but unnatural or clunky |
Hover an underline to read the note (a short English explanation with the natural alternative). Clicking an underlined word opens the right panel with the dictionary entry and the write-up note stacked together.
Annotations are anchored by exact quotes from your message. A quoted span the model invents (one that does not appear verbatim in what you wrote) is dropped rather than guessed at — no underline is better than a wrong one. A message with nothing to flag gets no underlines.
Clickable words
Every message — yours and the partner's — runs through the same morphological pipeline as the reader: conjugated phrases group together, and clicking any Japanese word opens the right panel with its reading, dictionary entry, and conjugation summary. From there, Learn adds the word to your reviews, and Known / Ignore set its status, exactly as in the reader. See Reviews & SRS for what happens next.
Level calibration
The partner's Japanese is calibrated from three signals:
- Recorded vocabulary — your known-word count maps to a rough JLPT band that seeds the prompt.
- Your own writing — the model is told this estimate may lag reality and that your actual messages are the better signal, so a small recorded vocabulary never caps the conversation.
- The challenge dial — a dropdown next to the input box: Match my level, Push me a little (the default), or Full immersion (natural native Japanese, no simplification). Changing it saves immediately and applies from your next message.
Conversations
The left sidebar lists past conversations; hovering shows the full title, message count, and start date. Conversations persist in the database — reopen one to continue it, or delete it with the trash button. New conversation starts a fresh thread.
In the input box, Enter sends and Shift+Enter inserts a newline.
Privacy
Only the text needed for the feature leaves the app, and only to the provider you configured: the sentence being explained, or the conversation history plus the level hint described above. Nothing else — no library contents, no statistics, no review data — is ever sent. With Ollama (or a local custom endpoint such as LM Studio), nothing leaves your machine at all.