Training is the agent’s knowledge base. While the prompt defines behavior, training provides the content — policies, catalogs, manuals, FAQs, and any information the agent needs to respond accurately.
Overview
Access the Training tab in the agent settings. You can add knowledge from five different sources, and they coexist in the same base.
How RAG works
Timely.ai uses Retrieval-Augmented Generation (RAG) to inject relevant knowledge into each execution:
- All added content is divided into chunks (excerpts) and transformed into embedding vectors.
- When the customer sends a message, the system searches for the chunks semantically closest to the question.
- The retrieved chunks are injected into the model’s context alongside the prompt and conversation history.
- The model generates the response based on this enriched context.
Only the most relevant excerpts are injected — not the entire document. This saves tokens and keeps the context focused.
Training types
Document
Website
Text
Q&A
YouTube
Upload files directly to the platform. Supported formats:
- PDF (manuals, contracts, catalogs)
- DOCX (Word documents)
- TXT (plain text)
- CSV (data tables)
Go to the Documents tab
In the training section, click Documents and then Add document.
Upload the file
Drag the file or use the file selector. Large files may take a few seconds to process.
Wait for processing
The system extracts the text, divides it into chunks, and generates the embeddings. The status changes to Processed when complete.
Documents with many images and little text (e.g., presentations with visual slides) have reduced extraction quality. Prefer exporting to PDF with selectable text.
Index web pages directly by URL. The crawler accesses the page, extracts the textual content, and processes it as training.Add the URL
Click Add site and enter the full URL (e.g., https://yoursite.com/privacy-policy).
Configure crawling
Choose whether the crawler should follow internal links on the page (deep crawling) or process only the entered URL.
Start crawling
Click Crawl. Progress is displayed in real time. Already-processed URLs appear in the list with the indexing date.
Pages protected by login or CAPTCHA cannot be crawled. For this content, use the Text or Document type.
For content that does not exist in a file or URL — internal policies, specific instructions, sales scripts.Open the text editor
Click Add text. A free text editor is displayed.
Paste or write the content
Add the text directly. You can format with Markdown for better structure.
Save and process
Click Save. The text is immediately added to the knowledge base.
Use headings (# Section) and lists to structure the content. The RAG system respects the structure and groups chunks by section, improving retrieval accuracy.
Question and answer pairs — the most accurate way to train an agent for specific frequently asked questions.Access the Q&A manager
Click Q&A in the training tab.
Add a pair
Click New question. Fill in the question exactly as a customer would ask it and the ideal response the agent should give.
Organize by category and tags
Use categories (e.g., “Shipping”, “Payment”) and tags to organize and facilitate searching in the manager. This does not affect RAG behavior, but helps with maintenance.
Add variations of the same question as separate entries with the same answer. This improves semantic coverage and retrieval rate for similar questions.
Transcribe YouTube videos and add the content to the knowledge base. Ideal for video training, tutorials, and webinars.Paste the video URL
Go to YouTube in the training section and enter the full video URL.
Validate the video
The system displays the title, channel, and thumbnail for confirmation. Click Process.
Wait for transcription
The system extracts the automatic YouTube transcript (or closed captions when available), divides it into chunks, and generates the embeddings.
Videos without automatic or closed captions on YouTube cannot be processed. The video status is displayed in the list after the processing attempt.
Knowledge base viewer
In the Training tab, you can open the Knowledge Base Viewer to:
- List all indexed chunks and their metadata
- Test semantic search — simulate a question and see which chunks would be retrieved
- Identify duplicate or outdated content for removal
Token limits
The volume of content that can be stored and processed per execution has limits depending on the plan:
| Plan | Total storage | Chunks injected per execution |
|---|
| Starter | 5 MB of processed text | Up to 5 chunks |
| Pro | 50 MB of processed text | Up to 10 chunks |
| Enterprise | Unlimited | Up to 20 chunks |
Adding more content does not guarantee better responses. The system injects only the most relevant chunks — the quality of content organization and writing is more important than volume.
Best practices
- Keep each document focused on a single topic. Overly extensive and generic documents reduce retrieval accuracy.
- Update training whenever policies or products change. Outdated content causes incorrect responses.
- Prefer Q&A for critical questions where the answer needs to be exact and consistent.
- Use the Knowledge Base Viewer to test the search before publishing the agent.
Next steps
Connect tools
Add tools for the agent to act beyond responding — scheduling, creating records, sending emails.
Test the agent
Use the Playground and semantic search to validate the knowledge base.
