ask

Ask OpenFeather: overview

How Ask turns a question and your authorised context into a grounded answer, what it can and cannot read, and which controls change the result.

· OpenbirdAI documentation

Ask is the OpenFeather conversation surface. You write a question in plain language, OpenFeather plans the request, reads only the sources the current account and workspace are allowed to use, and answers with citations. Ask is available to any signed-in account; what it can read depends entirely on what you have connected.

What happens when you send a message

  1. Your request is classified — conversational, private-data, web-research, location, product-help or a refusal-class request.
  2. A deterministic planner decides which sources are required and reserves the credit cost before any provider call is made.
  3. Authorised retrieval runs: connected integrations, workspace memory, and web research only when the request needs current external information.
  4. OpenFeather synthesises an answer and streams it to the page.
  5. Usage is settled against the reservation, and the conversation is persisted with its sources.
Credits are reserved before the work starts and released if the request fails. A failed run is refunded rather than charged.

What Ask can read

  • Connected integrations, limited to the services inside each authorised connection.
  • The current conversation, plus scoped memory for the active workspace and account.
  • Public web results, only when the request needs current or external information.
  • Authorised workspace records such as team context, goals and authorised analytics.

Ask never reads another user, another workspace, or provider credentials, and it never queries the database directly.

A product-help question uses public first-party documentation rather than scanning connected services. Advice can use information you supply without a connected account; generating an AI answer still calls a model provider.

Planning in the conversation

You can request an inline WorkMap: a proposed task plan with owners, dependencies, milestones and budget constraints. For example, ask for a two-week plan using an explicitly stated one-time budget. Your current request supplies the constraints; an older saved budget does not authorize new spending.

WorkMap generation is subject to validation and may return a clearly labeled incomplete result or offer a retry. A plan is not proof that its tasks were executed. Connected reads require the relevant connection and permission; supported actions can require approval. Documentation describes intended product behavior, not signed-in runtime acceptance for every feature or account.

The controls that change the result

  • Model / thinking mode — Core, Pro and Deep Strategy change answer depth and credit cost. See Credits and usage.
  • Plugin or channel selector — biases the request toward a chosen connected source.
  • New conversation — clears conversation context so the next question starts fresh.
  • Stop control — available while a response streams; it cancels the run and marks the turn cancelled.

A good first prompt

Specific questions produce specific answers. Compare:

text
"Tell me about marketing"          vague — OpenFeather must ask what you mean
"Summarise my Q3 sales by channel"     specific, answerable, verifiable
"What changed in this account's spend since June?"   private data, needs a connection

Limitations worth knowing

  • An empty answer does not mean the source has no data. It usually means the request was not authorised, the account was wrong, or the date range excluded it.
  • Generated analysis is interpretation, not evidence. Check citations for anything important.
  • Some capabilities depend on the provider, the granted permissions and your plan.

Common questions

Does Ask need an integration to answer questions?

General questions work with no connection. Questions about your own data require the matching integration.

Why does the same question give a different answer later?

Because connected data changes and because retrieval depends on the time range and the authorised context at request time. Check the sources listed with the answer.

Was this page helpful?