Home › Studio › Ask › Ask

Ask

/ask is the chat surface for asking metadata questions. It reads from the DB profile catalog, the document RAG index, and the codebase RAG index, and threads its answers with tool-call traces so you can audit how an answer was reached.

Sessions

The left sidebar (above the chat on mobile, side-by-side on md: and up) holds the Sessions list.

• + New session creates a fresh chat with the default scope.
• Each existing session row shows a title (auto-named from the first question, or manually renamed), the last-active timestamp, and a delete button.
• Click a row to resume the conversation. The agent replays the prior 4 question/answer turns into its context so follow-ups like "the second one" resolve without re-explaining.
• An End session control closes the current chat (it stays in the list, but the agent forgets in-flight state).

Scope dropdown

The selector above the textarea lets you pick which DB profiles /ask queries:

• Multi-select checkboxes (one per saved profile)
• All profiles option (default for a new session)
• Per-session sticky — the selection persists until you change it
• Focus auto-detection — when ≥60 % of the last three turns mentioned the same profile, Studio shows a Focus: <profile> (auto) hint to indicate the system prompt is biased toward it. The biasing doesn't lock out cross-profile questions; it just shapes the default answer when the question is ambiguous.

A small 📄 N docs · 💻 M code badge above the dropdown shows how many doc / code profiles are in scope for the active DB selection — useful when /ask answers seem under-grounded and you're not sure whether the right RAG profiles are linked.

Cache-only by default · Live refresh on demand

Every new session starts in cache-only mode. The agent answers from the local catalog cache, the doc RAG index, and the code RAG index without opening a live database connection — fast, free of warehouse cost, and safe to run anywhere even when the production DB isn't reachable.

When a question genuinely needs fresh data (recent inserts, a schema change, row-level numbers), the agent does not silently dial out. It returns a structured needs_live_refresh envelope and the UI surfaces a one-click "Enable Live refresh & retry" affordance under the assistant turn. Click it once to:

1. Flip the current session to allow live tool dispatches.
2. Replay the question with live retrieval enabled.

The toggle is per-session, so a different session in the same Ask tab stays cache-only. The composer also exposes a Live refresh switch you can flip preemptively if you know the next question needs fresh data; this becomes the new default for that session until you switch it back off.

Chat bubbles

• User turn — left-aligned bubble, light background.
• Assistant turn — right-aligned, Markdown-rendered with GFM tables and code blocks.
• Cancelled — thin pill marker on the turn the user interrupted.

Tool-call trace

Every assistant turn that called tools renders an expandable trace under the bubble:

• Tool name + collapsed JSON arguments
• Tool-call latency in milliseconds
• Result preview (truncated to a few lines)
• Citations, when present, as a collapsible block

While the agent is still streaming, the trace doubles as a real-time activity feed — each tool dispatch lights up the moment it fires, the per-tool timer starts ticking immediately, and a one-second LLM heartbeat shows up while the agent is between tool calls. A long tool round no longer looks like a hang.

The tool surface is split into two tiers:

Cache-only tools (run on every question without hitting the live DB):

Tool	Purpose
list_schemas	Schema list for one or more profiles.
list_tables_in_schema	Tables in a given schema.
describe_table	Cached column list + comments for a table.
describe_column	One column's full cached metadata — type, comment, profiling stats.
find_table_by_name	Resolve a short name to a fully-qualified address.
search_tables_by_concept / search_columns_by_concept	Semantic search over the catalog embeddings.
search_docs / search_code	Hybrid RAG retrieval over linked doc and code profiles.
get_join_candidates / find_joinable_tables / find_joinable_across_profiles	Suggest joins from FKs, view DDL, query-log co-occurrence, and column-name overlap.
lineage_for_table / lineage_for_column	Cached lineage neighbours, both manual and AI-generated.
catalog_coverage_summary	Per-profile description coverage % across tables and columns.
catalog_inventory	Counts of profiles, schemas, tables, columns, lineage artifacts, ingested code assets.
catalog_sync_status	One-call freshness report across every profile — answers "are my tables fresh?" with zero DB queries.
compare_runs / describe_run	History-store reads for /run audit questions.

Live tools (only available when Live refresh is on, surfaced through the needs_live_refresh envelope when needed):

Tool	Purpose
list_volumes	Live volume listing (Databricks Unity Catalog).
run_count_query / run_sample_query	Row-count or sample preview against the live DB.
refresh_schema_cache	Probe the live catalog and update AMX's cache row.

Asking about variations and mode

describe_run returns each result row with alternatives_mode at the top level and a variations[] array of v2 / v3+ descendants for that asset. Each variation entry carries its own mode (semantic / lexical), the seed_alternative_text the user picked, the descendant_run_id, and the variation's own alternatives. The agent is taught the semantic vs lexical contract: semantic means paraphrase of the seed (same factual content, different surface form); lexical means re-use of the seed's vocabulary with a distinct candidate meaning a reviewer can tell apart. Example prompts:

• "What variations did we try for public.country.abbreviation in run #99 — were they semantic or lexical?"
• "Evaluate the lexical variations on sales.orders.status in run #142 and name the distinct candidate meaning each version proposes."
• "Compare v1 and v2 of users.email_status in run #67 — what changed and what does the mode imply about the agent's reasoning?"

For Re-Run descendants the seed_alternative_text is null (Re-Run regenerates from scratch rather than anchoring on a seed); the agent still surfaces kind: "rerun" and the descendant's mode so you can reason about diversity decisions across versions. See Variations and Alternatives diversity mode.

Answer footer

Below each assistant turn:

N profiles · X.Ys · focus: WAREHOUSE · 1 234 tokens · $0.0042

Profile count, total latency, current focus, total token count, and the per-turn cost. The cost is computed at the rates active when the turn ran, not at current rates.

Message input

• Textarea with markdown syntax highlighting
• Submit with Ctrl/Cmd-Enter (or the Send button)
• A Cancel button replaces Send while the agent is streaming — pressing it sets the cancel token; the answer stops at the next tool-call boundary

A state.seedPrompt hand-off from other pages (e.g. the Compare page's Ask AMX button) pre-fills the textarea and auto-submits.

Cross-references

Studio	CLI
Ask chat	/ask
Sessions list	/session list
Resume session	/session resume <id>
End session	/session end
Scope dropdown	/session scope [profiles] and /use-db <profile> [<profile>…]
Tool-call trace	CLI prints tool names inline in the answer