Skip to content
Home › Studio › Operations › Settings

Settings

/settings is the profile-management surface in Studio. It mirrors the CLI's /add-…-profile wizards in a five-tab tabbed interface.

Database

Lists every configured DB profile. Each row shows the profile name, backend, host, and the database / catalog the profile pins.

Per-profile actions:

  • Edit — re-open the wizard pre-filled with current values
  • Delete — remove the profile (after confirmation)
  • Test connection — round-trips a small probe and renders a green check or red error message in a modal

The Add new profile wizard is a multi-step dialog:

  • Backend dropdown — PostgreSQL, MySQL, Snowflake, Databricks, BigQuery, Oracle, SQL Server, Redshift, ClickHouse, DuckDB
  • Dynamic form fields per backend (host / port / username / password / database / catalog / schema / SSL settings / proxy / timeout / …)
  • Inline help text per field
  • Test connection before save

LLM

Lists every configured LLM profile with provider, model, and the live per-1M-token price (auto-detected from the LiteLLM + OpenRouter pricing cache, surfaced in Pricing).

The Add new profile wizard:

  • Provider dropdown — OpenAI, Anthropic, Gemini, Databricks Serving, OpenRouter, DeepSeek, Ollama, local-via-LiteLLM, Kimi (see LLM Providers)
  • Model dropdown — populated from the provider's offerings, or a free-text field
  • API-key / base-URL fields appropriate for the provider
  • Alternatives per column slider (1–5) — how many candidate descriptions the LLM emits per column.
  • Alternatives diversity mode segmented control — choose whether those candidates are paraphrases (semantic) or shared-vocabulary candidate meanings (lexical). Disabled when alternatives per column is 1. See Alternatives mode for the full reference and worked examples.
  • Confidence signal dropdown — which per-alternative scorer drives the HIGH / MED / LOW pill on each row: self_consistency (default), logprob, self_decl, judge, none. See Confidence signals for what each one measures, score ranges, and how to tune the bands.
  • Optional per-model price override — pre-filled with the auto-detected public rate, so an override is a deliberate tweak rather than a re-keying chore

Docs

Lists configured doc profiles. Each row shows the profile name, source paths, and the DB profiles it's linked to.

The wizard exposes:

  • Name field
  • File path picker (filesystem or pasted)
  • Linked DB profiles multi-select (optional — empty = global)

On save, AMX kicks off async indexing in the background. The status badge on the row updates as the work progresses: Indexing → Ready → Failed. Indexing status can also be inspected from System → Catalog.

Code

Same shape as Docs, scoped to source-code paths. One path per profile (filesystem or Git URL — cloned on first scan into ~/.amx/code_cache/<slug>/).

Linked DB profiles control which /ask scopes consult this code profile.

Embeddings

Docs RAG and code RAG carry independent embedding providers (cfg.embedding_docs and cfg.embedding_code), so a code-specialised encoder can drive code retrieval while a prose-tuned model drives the documentation side. The tab renders two panels — one per side — sharing the same form shape:

  • Provider radio — MiniLM (Chroma bundled, offline default), OpenAI-compatible (OpenAI, OpenRouter, Together, Mistral, DeepInfra, Azure, vLLM / LM Studio / llama.cpp via base_url), or local sentence-transformers. The local option is auto-disabled with a hint when the host lacks the sentence-transformers extra.
  • Preset dropdown (OpenAI-compatible only) — fills base_url with the matching endpoint.
  • Model, Base URL, API key — rendered conditionally on the chosen kind. Keys are masked on read (********); leaving the placeholder in the save body preserves the existing secret.
  • Test connection — embeds a one-token sentinel with the merged- but-not-yet-saved settings and reports a green pill with the vector dimension, or a red pill with the provider's error verbatim.
  • Save — persists to cfg.embedding_{docs,code} and re-installs the process-wide embedding factory so subsequent /search and code-RAG queries pick up the new provider without restarting the server.

When the active provider drifts from what an existing collection was embedded with, a yellow banner appears with the concrete collection and chunk counts and a Rebuild button. The rebuild clears the affected vector store so the next ingest or query re-embeds with the active provider; document and code data themselves are untouched.

Profile linking

Both Docs and Code wizards expose a linked-DB-profiles multi-select. This controls which /ask invocations see the profile's contents — a doc profile linked to prod-pg is silent when /ask is running against dev-snowflake, unless the link list is empty (= global).

CLI equivalents

Studio CLI
DB tab /add-db-profile, /edit-db-profile, /remove-db-profile, /connect
LLM tab /add-llm-profile, /use-llm, /temperature, /max-tokens, /n-alternatives, /cost
Docs tab /add-doc-profile, /doc-link, /scan, /ingest
Code tab /add-code-profile, /code-link, /code-scan
Embeddings tab /embeddings docs, /embeddings code