Train agents with a knowledge base

Wassla agents answer customers by reading from a knowledge base you fill yourself. You add knowledge in three ways — upload documents, paste a public URL, or write entries by hand — and Wassla converts each one into searchable chunks so the agent can quote the right passage on demand. This page walks through all three flows and explains, in plain English, how the retrieval works under the hood.

Open the Knowledge page

  1. In the workspace sidebar, click Knowledge. The page header reads "Everything your agents know — synced, versioned, and always cited."
  2. If you have no knowledge base yet, click Create knowledge base in the empty state and give it a name (for example Support, Pricing, or Returns).
  3. Use the KB switcher at the top of the page to pick which knowledge base you are editing. Each agent can be scoped to one or more knowledge bases later.

You will see four tiles on the Add what you know row: Upload a document, Paste a link, Write it yourself, and Connect a tool. Pick whichever matches the source you are bringing in.

Upload documents (PDF, Word, images, text)

Use this when your knowledge already lives in a file — a product PDF, an internal Word doc, a screenshot of a return policy, or a Markdown export.

  1. On the Knowledge page, click the Upload a document tile, or drag a file anywhere on the page.
  2. Accepted formats: PDF, DOCX, TXT, Markdown, and image files (PNG, JPG, JPEG, WEBP). Maximum size is 12 MB per file.
  3. Wassla runs the file through OCR (for scanned PDFs and images) and an LLM structurer that splits the content into proposed knowledge entries with titles and categories.
  4. The Review step is non-negotiable — you see every extracted entry before anything is saved. Untick anything that is pricing, customer data, or marketing copy you do not want the agent to quote.
  5. Click Confirm. Selected entries are inserted into the knowledge base and start indexing immediately.

If you ever upload the wrong file, open any entry from the Sources table and delete it; the agent stops quoting it within seconds.

Paste a URL (single page or whole-site crawl)

Use this when your knowledge lives on a public web page — your help center, a policy page, or your product site.

  1. Click the Paste a link tile. If you have not picked a knowledge base yet, the KB switcher flashes red so you know where to start.
  2. Enter the URL. You can toggle Crawl linked pages on the same domain if you want Wassla to follow internal links and pull a multi-page set instead of just the one page.
  3. The extractor fetches the page, strips navigation and scripts, and asks an LLM to organise what it finds into entries — FAQs, contact info, social links, menu items, features, and pricing. A typical import takes 10 to 90 seconds; a rotating status line shows you what step is running.
  4. The Review step shows you each candidate entry with a checkbox. Tick the ones worth keeping and click Confirm.

URL imports authenticate against your tenant membership and block attempts to reach internal IPs or cloud metadata endpoints, so the importer cannot be tricked into pulling private infrastructure.

Write entries yourself

Use this for the answers only your team knows by heart — internal SLAs, edge-case policies, or a phrase you want the agent to use verbatim.

  1. Click the Write it yourself tile. The New answer modal opens.
  2. Give the entry a clear title (this is what shows up in the Sources table) and write the body in plain text or Markdown.
  3. Optionally assign a category and subcategory so you can group related answers.
  4. Click Save. The entry is embedded and searchable within seconds.

When an agent gets a question wrong, the fastest fix is almost always to add one short, hand-written entry that says the right thing in the words your customers actually use.

How retrieval works, in plain English

When a customer asks a question, the agent does not "search keywords" the way Google does. It does something called retrieval-augmented generation, or RAG. Here is what happens:

  1. Wassla converts every knowledge entry into an embedding — a list of 1,536 numbers (generated by OpenAI's text-embedding-3-small model) that captures the meaning of the text. Two passages about "refunds" land near each other in that number-space even if one says "money back" and the other says "reimbursement."
  2. When a question comes in, Wassla generates an embedding for the question the same way.
  3. Wassla then asks Postgres (using the pgvector extension) to find the entries whose embeddings sit closest to the question's embedding. That is the semantic similarity match.
  4. To catch exact-name and rare-term questions that semantic search can miss, Wassla runs a keyword pass over titles and content in parallel.
  5. The two result sets are merged, deduplicated, and the top matches are handed to the LLM, which writes the customer-facing reply and shows the source as a citation.

Repeat questions hit a 24-hour embedding cache, so the same question from a hundred different customers only pays for one embedding round-trip.

Verify the agent is using your knowledge

  1. Open the test inbox from the agent page and ask the kinds of questions a real customer would.
  2. Every answer shows its sources next to it. Click a source to jump straight to the entry it quoted.
  3. If the citation is wrong, edit the underlying entry (or add a new one) and re-test. Wassla re-indexes on save.
  4. Check the Knowledge gaps panel on the Knowledge page — it lists real questions customers asked that your agents could not answer well. Each gap is a one-click prompt to write the missing answer.

Indexing freshness and limits

  • Hand-written entries and document uploads are indexed within seconds.
  • URL sources can be re-imported on demand from the source card.
  • Each tenant gets up to 120 knowledge searches per minute before rate-limiting kicks in — high enough for a busy inbox, low enough to stop a runaway loop.
  • The Sources table is sorted by most-recently-updated and shows a Syncing chip while embeddings are still generating, then flips to Synced when the entry is searchable.