Nia documentation · 4.0.0
BrainDump
Configure optional voice capture, mixed task/shopping suggestions and recipe text/refinement with your model providers.BrainDump is optional and disabled by default. It proposes content for the user to review; it does not silently save a recording or an imported recipe as finished content.
Three workflows, one account permission
- Tasks and Shopping: record a thought, transcribe it and review suggestions. One recording can contain both tasks and shopping items. Each suggestion has an editable module and destination; unavailable or ambiguous destinations must be corrected before saving.
- Recipe text import: paste recipe text and let BrainDump structure an unsaved recipe draft.
- Recipe refinement: open a manual or website-imported draft and choose Revise with BrainDump. BrainDump can clean the description, improve step headings, suggest names for unnamed ingredient groups and link ingredients to preparation steps.
Recipe link import is a separate, non-AI feature. It extracts supported public recipe pages using recipe-scrapers and structured recipe data. It does not require the BrainDump permission. Both link import and BrainDump need a server connection; users save accepted drafts through the ordinary recipe editor.
What you need
Voice capture uses two provider stages:
- Speech to text (STT): a local or remote whisper.cpp-style transcription endpoint.
- Language model (LLM): an OpenAI-compatible endpoint or Ollama.
Recipe text import and refinement use only the LLM stage; they do not require working STT or microphone access. The core Tasks, Shopping and Recipes modules, personal destination memory and recipe link import work without an AI model.
Supported LLM setups include local services such as LM Studio or Ollama, Ollama Cloud and OpenAI-compatible gateways. Compatibility depends on the endpoint implementing the APIs expected by the selected mode.
Configure BrainDump
In /admin:
- Configure the LLM provider, base URL, model, optional API key and timeout.
- For voice capture, configure the STT URL, optional token, language and timeout.
- Use the built-in connection tests for the stages you need.
- Enable BrainDump globally.
- Enable BrainDump for each user who should have access.
Global enablement alone does not grant access to every account. The same per-user gate applies to recipe text import and refinement. Turning off BrainDump does not disable non-AI link import or personal destination memory.
Prompts and recipe review
The admin panel shows the active Nia default extraction prompt from the backend. Start with Use Nia default. Custom instructions can be appended, or can replace the extraction prompt; test them with mixed Tasks/Shopping input, explicit destinations and ambiguous dates before normal use.
Recipe BrainDump uses its own recipe-specific structure and refinement instructions with the same configured LLM connection. The Tasks/Shopping extraction-prompt override is not a recipe prompt editor. Refinement preserves manual ingredient choices and explicit exclusions, and distinguishes similarly named ingredients in different groups. It leaves uncertain step quantities open rather than inventing missing amounts. Users should check portions, ingredient amounts, step assignments and any specific issues before saving.
A suitable model still matters: connection-test success does not prove that the model follows the recipe schema or reliably reasons about partial quantities. Test a real multi-group recipe as well as a simple task recording. A rejected or timed-out refinement leaves the existing editor draft available for correction or retry.
Network placement and imports
Provider URLs are requested by the Nia server. A local provider may use an internal hostname reachable from the container or service host. In Docker, 127.0.0.1 refers to the Nia container itself; use a shared Docker network, service name or explicitly routed host address.
Recipe website import has a different network policy: it accepts public HTTP/HTTPS pages on standard ports and checks redirects and resolved addresses. Private-network URLs, credential-bearing links and login-only pages are not supported. Do not disable these protections to import an internal page; copy its recipe text instead if appropriate.
Privacy and limits
Voice audio goes to the configured STT service; transcripts and bounded destination metadata go to the LLM for extraction. Recipe text or the structured draft goes to the LLM for recipe work. Recipe images are not sent to that model. Local providers keep processing on your infrastructure; remote providers receive the submitted content under their own policies. Link import separately contacts the recipe website and, when available, the source image host.
Destination memory stores personal assignment history, not recordings or transcripts. Its learning/reset controls are independent of BrainDump provider settings.
Set realistic timeouts and test the languages and recording lengths your users will submit. Recipe refinement can need more than one model pass within one total deadline, capped at 100 seconds even if a longer LLM timeout is configured. Import and recipe BrainDump requests have bounded input/output, concurrency and per-account rate limits; repeated requests can temporarily return a busy/rate-limit message.
Microphone availability depends on HTTPS, platform permissions and the client. Text-based recipe import does not need microphone permission.
Implementation references: shared AI access gate, extraction configuration, recipe refinement and public import fetch policy.