Skip to main content
The most common knowledge source is the document — manuals, contracts, policies, product catalogs, or any file you already have on hand. Timely.ai extracts the text, splits it into chunks, and generates embeddings automatically after upload.
Knowledge Base document list

​
Uploading Documents

1

Access the agent's Knowledge Base

Open the desired agent and click the Knowledge Base tab, then select Documents.
2

Click Add document

Click the upload button or drag the file directly into the designated area.
3

Wait for processing

The file is sent to Supabase Storage and the knowledge-process-document job is triggered automatically. The document status changes from processing to ready when complete.
4

Verify the result

Click the preview icon to review the extracted text, the automatically generated summary, and the identified keywords.

​
Formats and Limits

ParameterValue
Accepted formatsPDF, TXT
Maximum file sizeDefined by the account plan
Target size per chunk~4,000 characters (~1,000 tokens)
Overlap between chunks~400 characters (~100 tokens)
Maximum embedding batch100 chunks per request to OpenAI
Maximum characters for metadata8,000 characters of content
During processing, the document has processing status and does not yet appear in semantic searches. Wait for the status to change to ready before testing the agent.

​
What Happens During Processing

The internal pipeline executes the following steps:
  1. File reading: the content is read from Supabase Storage.
  2. Text extraction: for PDFs, text is extracted page by page.
  3. Intelligent chunking: the text is split prioritizing paragraph breaks (\n\n). Paragraphs that are too large are subdivided by sentences (". "). If still too large, forced character-level splitting is applied.
  4. Metadata generation: name (up to 80 characters), description (up to 250 characters), and keywords are generated by the LLM from the first 8,000 characters.
  5. Embeddings: each chunk is sent to OpenAI’s embeddings API in batches of up to 100.
  6. Storage: chunks and vectors are saved in the knowledge_chunks table.

​
Managing Documents

ActionDescription
ViewOpens the extracted content and metadata of the document
DownloadDownloads the original file from Storage
DisableRemoves the document from searches without deleting the file
DeleteRemoves the document, all chunks, and the file from Storage
Deleting a document is irreversible. All associated chunks and embeddings are permanently deleted. Download the original file before deleting if you need to keep it.

​
Best Practices for Documents

Prefer selectable text

PDFs scanned as images have no extractable text. Use PDFs with native text or convert them before uploading.

Split large documents

Documents with hundreds of pages generate many chunks. Splitting by section improves semantic search precision.

Keep content up to date

Delete outdated documents and upload the new version. The agent may retrieve old information if the outdated document is still active.

Name files clearly

The filename and generated metadata are used in search results. A descriptive name improves traceability.
After indexing a new document, use the integrated semantic search in the Knowledge Base to test whether relevant questions return the correct excerpts before activating the agent in production.