Build a knowledge base

An AI assistant only knows what you tell it. A knowledge base is your own material (a price list, a treatment FAQ, a returns policy, your website) grouped into collections that the assistant can search while it writes a reply. Add the price list once, and every AI Reply step pointed at that collection quotes the current prices. It lives at Settings → AI Knowledge, and it is workspace-wide: build a collection once and any automation in the workspace can search it. Prefer to watch first? This walkthrough covers uploading your business data for the chatbot to use.

​

The two knowledge screens

DMLY has two places that show the same underlying collections, and it is worth knowing which one you are on. Use Settings → AI Knowledge. The Vector stores tab is the older per-provider panel; it is still there, and when your workspace replies on that provider’s key the two screens list the same collections. Everything below describes the Settings screen unless it says otherwise; the provider tab is covered in Vector stores on a provider panel at the end.

​

Before you start

• The AI agent is on every plan. There is nothing to unlock before you open AI Knowledge.
• It works before you connect anything. Knowledge runs on whichever AI answers for your workspace: your own connected OpenAI or Google Gemini key if your plan allows one and you have connected it, otherwise DMLY’s built-in AI. A line at the top of the page names the account it is backed by, and confirms the knowledge is private to your workspace.
• Claude and DeepSeek cannot search documents. If an AI Reply step replies on one of those, it gets no knowledge whatever you have built. Use OpenAI or Gemini for a step that needs to read your material.
• Teammates need the right permission. Viewing the page needs access to the Setup & Integrations module. Creating or deleting a collection, adding, importing or removing a source and running the test box all need Manage bot setup. See Roles and permissions.

​

Build a collection

​

The four source types

Add source offers exactly four kinds, listed under Source type: The list shows each row’s type as File, URL, FAQ or Text, with the page address underneath for anything ingested from the web. There is no in-place editing in the sources list. To change an FAQ or a pasted note, select Remove on the row and add it again.

​

Import a whole site at once

Adding a documentation site one page at a time is a job people abandon halfway, which leaves the assistant confidently half-informed. Import pages on a collection does it in bulk. The confirmation reports every address’s fate, not just the wins: how many pages were imported, how many were already in the collection, and how many failed (with the first failing address named). A page already in the collection is skipped rather than added twice, so re-running an import is safe.

​

Find a source in a big collection

Each collection’s header shows how many sources it holds and how many automations use it, and the list itself is paged:

• 20 sources per page, newest first, with Previous and Next and a page count.
• A Search sources by name or address… box appears once a collection holds more than 20 sources. It searches the whole collection, not just the rows on screen, and matches both the source name and its web address.
• Remove on a row deletes that source from the collection and from the AI account behind it.

Delete collection removes the collection and every source in it, and cannot be undone.

​

Indexing status

Adding a source hands it to the AI account, which indexes it in the background, so a new row starts at processing and settles by itself:

• processing: still indexing. Nothing to press.
• ready: searchable.
• error: indexing failed, and the reason is shown on the row.

The page updates the statuses on its own only if a source was already indexing when you opened it. After you add or import sources, reload the page to see where they got to. Indexing is also reconciled in the background every few minutes, so a source settles whether or not anyone is watching the page.

​

Test what it will find

The Test knowledge panel appears once you have at least one collection: Ask a question to check what the assistant would retrieve.

1. Leave the scope on All collections, or pick one collection to test it alone.
2. Type the question and select Test.
3. You get the answer, plus Retrieved snippets naming the source each snippet came from.

The test is deliberately strict: it answers only from the documents it retrieved and says it couldn’t find the answer otherwise. So “couldn’t find it” is a real result, and it means the material isn’t there (or isn’t phrased the way the question is). With no sources yet you get Add some sources first, then test. Nothing here reaches a contact and nothing is written to the Inbox.

​

Point an automation at it

Building a collection doesn’t make an assistant use it. On the AI Reply step’s Built-in Tools tab, the AI Knowledge block (with a Manage link back to this page) reads Let the assistant answer from your uploaded knowledge (files, URLs, FAQs). and offers four choices: All workspace knowledge is the right answer for most support assistants: a collection added next month is searched without editing the flow. Pick Selected collections when a sales bot and a support bot should read different material. Each collection’s line on the Settings page counts the automations that reference it, so you can spot a collection nothing reads from. Only published automations count, so a collection wired only into a draft still reads zero.

​

What makes the answers good

The assistant searches your collections and writes an answer from what it finds. So the quality of the answer is mostly the quality of what you put in. This helps:

• Write for the question, not for the page. A document that reads like an FAQ (a heading that is the question, a short answer under it) is far easier to search than a brochure. If contacts ask “do you take walk-ins?”, have a line that says exactly that. The FAQ (question & answer) source type exists for precisely this.
• Keep one subject per collection. Prices and Policies as two collections lets you point a sales bot at one and a support bot at the other.
• Say the specifics. Prices, opening hours, addresses, service names, how long something takes. These are what people actually ask for, and they are exactly what the assistant can’t invent.
• Keep one current version. Adding a new price list adds to the collection; it doesn’t overwrite the old one. Two versions in one collection means the assistant can find and quote either.

This hurts:

• A scanned or image-only PDF. If you can’t select the text with your cursor, neither can the search. Nothing errors; the source just contributes nothing.
• Marketing copy. Pages of tone with no facts give the assistant nothing to answer with.
• Contradictions. Two documents with different prices produce confidently wrong answers, and they’re the hardest kind to notice.
• Importing everything. A whole site imported wholesale, blog posts and press releases included, buries the price list among things nobody asked about. Import the pages that answer questions.

​

When it doesn’t answer from your material

The AI ignores your knowledge completely

Four causes, in the order worth checking. The step’s AI Knowledge selector is still on No knowledge, which is where a new step starts. Or the step replies on Claude or DeepSeek, which can’t search documents at all; switch it to OpenAI or Gemini. Or the collection is empty, or everything in it is still processing. Or the collections were built on a different AI account from the one now answering (see the warning above).

It answers, but with old prices

The old file is still in the collection. Adding a new one adds to the collection, it doesn’t replace what was there, so both versions are searchable. Remove the old row, or create a fresh collection with only the new file and tick that one on the step instead.

It found nothing in a PDF that clearly says it

Open the PDF and try to select the text. If you can’t, it’s a scan (an image of words, with no words in it). Re-export it from the original document, or retype the key facts as a Pasted text source.

An import brought in fewer pages than the site has

One import adds at most 100 pages. The confirmation says how many are still outstanding: run Import again with the same address and it carries on from where it stopped. If a sitemap returned nothing at all, check the address, or paste the page addresses instead.

The Tool calls trace looks empty

That’s expected, not a fault. Knowledge search runs inside the AI provider rather than in DMLY, so it never shows in the trace. An assistant answering perfectly from your collections still shows an empty trace.

The page says it needs a knowledge-capable AI

AI Knowledge needs a knowledge-capable AI (OpenAI or Gemini). Connect your own OpenAI or Google Gemini key under Integrations, and the page comes back with that account behind it. See Connect OpenAI.

The AI step didn't reply at all

Your plan is never the reason: the AI agent is on every plan. When a step is skipped it leaves a note in the conversation starting with a sparkle, and that note names the cause. The ones you are likely to meet:

• ✨ AI step skipped — AI replies are turned off for this workspace. The workspace kill switch Let the AI reply is off. Turn it back on under Conversation settings.
• ✨ AI step skipped — no AI provider is connected. Neither your own key nor DMLY’s built-in AI is available to answer.
• ✨ AI reply limit reached for this month — it resets at the start of next month. You have used your plan’s monthly AI reply allowance on DMLY’s built-in AI. The note goes on to point you at connecting your own key (replies on your own key are never counted) or at a bigger plan. See Plans.
• ✨ AI reply withheld — a teammate took over the conversation. Someone paused the bot while the model was still writing, so the answer was held back rather than sent over them.
• ✨ AI step skipped — per-conversation AI limit reached. A safety cap on how many times AI can answer inside a single run of a flow. Other conversations are unaffected.

​

Vector stores on a provider panel

The older per-provider panel is still available, and it is the only knowledge screen scoped to a single connected key. Go to Integrations, open OpenAI or Google Gemini, and select the Vector stores tab. It appears only for providers with native file search, and only once the provider is connected: until then it reads Connect the provider first to manage vector stores.

• Create a vector store takes a Store name and creates a collection on that provider’s account: Group documents the AI can search at reply time. Add a store, then upload PDFs / text files or ingest a web URL.
• Each store lists its files with an indexing dot, and offers + Upload file and an address box with Add URL. The note under the list reads PDF, TXT or MD up to 20 MB. URLs are fetched server-side and indexed.
• Refresh status re-checks indexing at the provider. Indexing… and Indexing failed show per file, with the reason under a failure.
• There is no way to remove a single file here: Delete applies to the whole store. Remove an individual source from Settings → AI Knowledge instead.

Stores created here are the same collections you manage in Settings → AI Knowledge, as long as that provider is the account your workspace replies on. What this tab cannot do is FAQs, pasted text, bulk import, search, the test box or the usage counts, which is why the Settings screen is the one to build in.