Knowledge Bases
Give a hosted Skysay agent private business documents it can search during a live call.
Knowledge Bases let a hosted phone agent answer questions using your own business material: opening hours, services, policies, menus, product details, FAQs, booking rules, or internal operating instructions.
The agent does not memorize the whole document or receive it on every call. When a caller asks a relevant question, the agent searches only the knowledge bases you attached to that agent and uses the matching passages to answer.
Available to every workspace
Knowledge Bases are included in the workspace. Open Knowledge in the dashboard to create your first base; no operator activation is required.
Set up a knowledge base
Create a base
Use one clear subject per base, such as Customer handbook, Sauna bookings,
or Product catalogue. Keeping unrelated material separate makes agent access
easier to understand and maintain.
Upload source documents
Upload a TXT, Markdown, PDF, or DOCX file up to 20 MB. Skysay stores the source privately, extracts its text asynchronously, and creates a searchable index. The document is usable when its status becomes Ready.
Choose which agent may use it
Under Agent access, attach the ready base to one or more hosted agents. An agent can search only bases explicitly attached to it. Detaching a base removes it from future calls without deleting the source. The agent's Knowledge section shows on which kinds of call the search runs; see What runs on calls.
Test with a real question
Call the agent and ask something answered by the document. Open the completed call in Calls to see the source titles used for the answer. Skysay does not expose the caller's search query or raw retrieved passages in the dashboard.
How retrieval works
During an authorized LiveKit call, the model can invoke one bounded tool:
search_knowledge_base(query="Are towels included with the sauna booking?")Skysay derives the workspace, agent, and allowed bases from the call. The model cannot choose another workspace, supply document identifiers, raise the result limit, or turn document text into another tool call. Retrieval returns at most five short, sourced passages.
Uploaded content is treated as untrusted data. Text inside a document cannot change the agent's permissions, call another tool, send a message, or override the agent's system instructions.
In a workflow
An agent with a published workflow does not offer its model this tool. Instead, a workflow Read tool step set to Search the knowledge base searches the bases attached to the agent, as in the Customer support and Appointment booking starter templates. Attaching a base is still what makes it searchable; a workflow never attaches one.
- The platform works out the question from what the caller said.
- The answer is never text a model wrote. A model only picks which sentences
of the matching passages answer the question; Skysay checks that they are
the document's exact words and quotes each whole sentence they come from, at
most 300 characters in all. So the answer is in the document's language. A
later line says it as
{{tool.knowledge_answer}}, inside its own section:{{#tool.knowledge_answer}}{{tool.knowledge_answer}}{{/tool.knowledge_answer}}. - When no base is attached, or nothing in the documents answers the question, the step takes its error route, so the workflow can say it does not know or hand the caller on.
- A call that holds a protected value runs no search.
Each search is recorded on the call like any other. The workflow builder's
Chat test does not search: once a base with a ready document is attached,
it fills the answer with a placeholder (test-knowledge_answer); before that,
the step takes its error route as on a call. See
Tools on a workflow call.
Updating and deleting content
- Upload a new source when business information changes.
- Retry a failed document after correcting the source or provider problem.
- Detach a base when one agent should stop using it.
- Delete a document or base to hide it from new retrieval immediately; private object and index cleanup then finishes asynchronously.
Privacy and processing
Source files are Customer Content. They are stored in Skysay private object storage, not published at a public URL. Google Gemini processes the source text and live search query needed to create embeddings and retrieve relevant passages. In a workflow, Google Gemini also picks which sentences of those passages answer the question; the answer the caller hears is the document's own words. A source is available only to the hosted agents you attach to it.
See the Privacy Policy for the wider data-processing terms.
Current V1 boundaries
The first release supports direct file uploads and hosted agents using Gemini. Website crawling, cloud-drive and CRM connectors, OCR, customer-configurable model providers, quotas, and knowledge-specific billing are planned follow-up capabilities rather than hidden behavior in this release.
For API automation, see the Knowledge Base routes in the
OpenAPI reference, including POST /v1/knowledge-bases, the
document upload routes, and PATCH /v1/agents/{agent_id}/knowledge-bases.