Knowledge in Creatio AI Studio
Knowledge sources let you connect prompt agents to your organization's data — internal documents, policies, product materials — so they can find and use the right information when responding to users. This increases the accuracy of agent responses and reduces the risk of the agent generating incorrect or made-up information.
Without knowledge sources, agents rely only on general data from their training. They do not automatically know your company's processes, guidelines, or internal content. Knowledge sources solve this by giving agents access to a defined set of your data to search at the moment a user asks a question.
Knowledge source retrieval process
When a user sends a message, the agent does not simply recall facts from its training data. Instead, it searches the attached knowledge sources for content that is relevant to the question, retrieves the most relevant fragments, and uses them to compose its response. This approach — searching a defined set of data at query time rather than relying on training — is called retrieval-augmented generation (RAG). It keeps responses grounded in your actual content and helps prevent the agent from generating inaccurate or made-up information.
To make retrieval efficient, Creatio AI Studio breaks source content into smaller chunks before indexing. The system then scores each chunk against the incoming query and passes the highest-scoring ones to the agent. You can improve this ranking by turning on Keyword boosting on a knowledge source — this gives extra weight to chunks that contain keywords from the query, which helps in cases where exact terminology matters.
Knowledge sources
A knowledge source is a named, indexed collection of files that an agent can search. Each source has a description that tells the agent what the source contains and when to use it. A precise description improves retrieval accuracy — the more specific it is, the better the agent can decide whether to query that source for a given request.
Creatio AI Studio offers two source types out of the box. File takes documents you upload in the following formats: *.pdf, *.txt, *.md, *.json, *.docx, *.pptx. Webpage takes a list of public URLs. A third type, Creatio data, reads live CRM records. It is enabled per organization and appears only where it has been turned on.
A knowledge source goes through the following statuses:
- "Draft" — the source has been created but has no content yet.
- "Indexing" — Creatio AI Studio is collecting and processing the content.
- "Ready" — the source is indexed and available for retrieval.
- "Failed" — fetching or indexing did not complete successfully.
- "Cancelled" — processing was stopped.
You can filter the Knowledge list by these statuses with the buttons above the list.
Each file of a File source also has its own status. It appears in the Extraction column of the Files tab on the source page. A file you upload reads "Pending" until processing starts, then "Indexing," and "Ready" when its content can be used. A file that reads "Failed" could not be processed, so its content is not available to agents. Remove the file and upload it again, or upload a different file.
Processing starts on its own: a source you create, or upload new files to, moves into "Indexing" without you asking it to reindex.
A source answers questions once it is "Ready." A webpage source keeps answering from its current content while it is re-fetched. Creatio AI Studio does not stop you from attaching a source that is still indexing or has failed, and the agent's picker lists those too, so check the status of every source you attach. An agent that answers from its training data instead of your content is often an agent attached to a source that never reached the "Ready" status. To check whether the agent searched a source during a run, open the run on the Agent Observability page: each knowledge source search appears in its Execution timeline. Learn more: Monitor agent runs and read traces.
Citations
When an agent uses a knowledge source to compose a response, it can include references to the source content that informed the response. This makes the agent's answers traceable — users can see where the information came from. Citations are configured per knowledge source at the agent level and can be turned on or off independently for each attached source.
Attaching knowledge sources to agents
A knowledge source does not become active on its own — it must be attached to a prompt agent. Each prompt agent has a Knowledge tab where you select which sources the agent can query. The picker attaches one source at a time, and you can attach as many as the agent needs, with citations configured separately for each one. The agent searches all attached sources when generating a response and combines the retrieved content as needed.
Files attached to a conversation
Not every document needs to become a knowledge source. A knowledge source is set up once, indexed, and attached to an agent, so every conversation with that agent can search it. A file that someone attaches to a conversation works right away, with no setup, and serves that conversation.
The agent can search text files and PDF, Word, Excel, and PowerPoint documents for a phrase and quote the matching passages. This way, it answers a question about one value or passage without reading the whole document. In a long file, the agent reads page by page and can use content from any part of it. Each time the agent lists, reads, or searches an attached file, the step appears in the run's Execution timeline on the Agent Observability page.
Use a knowledge source for content that many conversations need, and an attached file for content that matters to one conversation.