A voice agent is only as good as what it knows. The voice, the speed and the manners matter, but the moment a caller asks "Do you open on Sundays?" or "What does the deposit cover?", the only thing that counts is whether the agent has a clear, current answer to give. That answer comes from your knowledge base: the price lists, policies, menus and FAQs you hand to the agent.
Most businesses already have this material. The trouble is that it was written for people reading a page, not for an agent answering a question out loud in the middle of a conversation. A price list buried in a 30-page brochure, a policy that contradicts last year's version, or a FAQ that answers questions nobody asks will all produce hesitant or wrong answers.
This guide is for anyone setting up Cypher V, the Cypher LM voice agent, and for anyone who already has an agent live and wants it to answer better. We will cover what to include, how to structure it, which file formats work, how to bring documents in from Google Docs and Drive, how to test, and how to keep it all current.
Why the knowledge base matters more than the script
When people set up a voice agent for the first time, they tend to spend most of their effort on the greeting and the personality. Those are worth getting right, but they rarely decide whether a call goes well. Calls go well when the caller gets a specific, correct answer quickly, and calls go badly when the agent has to say it does not know something the business clearly should know.
A Cypher LM agent answers from two sources. The first is what it learns when you build it: the setup wizard reads your business website and collects your business details, services, audience and the way you want the agent to behave. The second is the knowledge base: the documents you upload or import. The wizard gives the agent a general picture of your business; the knowledge base gives it the detail.
Think of it the way you would think about onboarding a new receptionist. You would not hand them a script and hope for the best. You would give them the price sheet, the opening hours, the cancellation policy and a list of the questions customers ask most, and you would make sure those documents agree with each other. A voice agent needs exactly the same pack, written just as carefully.
The payoff is practical. A well-built knowledge base means fewer calls ending in "someone will call you back", fewer follow-ups for your team, and recaps that show the question was answered on the spot rather than deferred.
What to put in it: the six documents most businesses need
You do not need dozens of files. Most businesses get the bulk of the value from a short set of focused documents, each covering one subject well.
- Services and prices. Every service you offer, what it includes, how long it takes and what it costs. If prices vary, say how and why.
- Hours and locations. Opening hours by day, holiday closures, addresses, parking and how to find you.
- Policies. Booking, cancellation, refunds, deposits, late arrivals, warranties: anything a customer might push back on.
- Frequently asked questions. The questions your team answers every week, with the answer you actually give.
- Process guides. What happens after someone books, what to bring to a first appointment, how an order is delivered.
- Contact and escalation details. Which email or number to give for what, and what the agent should do when a question is out of scope.
The FAQ document deserves extra attention, because it is the closest match to how callers actually speak. Ask whoever answers your phone today to write down the twenty questions they hear most, in the caller's own words, with the answer they give. That one document often does more than everything else combined.
Leave out material that does not help a caller: internal process notes, staff rosters, marketing copy written to impress rather than inform. Anything in the knowledge base is something the agent may repeat, so only include what you are comfortable having said to a customer.
How to structure documents so an agent can use them
An agent reads your documents differently from a person skimming a page. It looks for the passage that best matches the caller's question and answers from it. That means the best documents are the ones where each answer can stand on its own.
One topic per section, one answer per question
Give every topic its own heading and keep related facts together. If the cancellation policy is split across three pages, with an exception in a footnote, the agent may find one part and miss the rest. Put the whole policy, including exceptions, in one place.
Write FAQs as a question followed directly by its answer. "Do you take walk-ins? Yes, Monday to Friday before 3 PM. On Saturdays we see booked appointments only." That format maps neatly onto what a caller asks.
Be specific and say it the way you would say it out loud
Vague wording produces vague answers. "Prices start from a competitive rate" gives the agent nothing to say. "A standard consultation is 45 minutes and costs the same for new and returning clients; see the price table for the current amount" gives it a real answer.
Remember that the answer will be spoken. Long tables with dozens of columns are hard to read aloud. Where you can, add a sentence that summarises the table: "Weekday appointments are cheaper than weekend ones, and the first visit includes an assessment."
Remove contradictions before you upload
If two documents disagree, the agent cannot know which one is right. The most common culprits are old price lists left next to new ones, and policies that changed without the FAQ being updated. Before you upload, check that each fact appears in one place, or that every copy matches.
Date your documents in the title or first line ("Price list, updated June 2026"). It helps you spot stale files later, and it gives the agent a cue when two passages look similar.
File formats, limits and what works best
Cypher LM accepts knowledge files in PDF, DOCX, CSV and TXT formats, up to 4 MB each and up to 10 files at a time. Each format has its strengths.
| Format | Best for | Watch out for |
|---|---|---|
| Brochures, policies, menus you already have as documents | Scanned images of text; image-heavy files that exceed 4 MB | |
| DOCX | FAQs and policies you edit often | Content hidden in text boxes or comments |
| CSV | Price lists, service catalogues, stock lists | Unclear column names; one fact split across columns |
| TXT | Short, plain answers you want read exactly | Missing structure; add headings in plain text |
If a file cannot be read, the dashboard tells you which one, and the usual fix is to save it again as a PDF. If a file is larger than 4 MB, split it into smaller files or remove images: the agent only reads the text, so pictures add size without adding answers.
Once a file is uploaded, you can preview it. The preview shows the text your agent reads, which is the quickest way to confirm that a table or a scanned page came through as real text rather than as nothing at all.
Shared knowledge versus agent knowledge
Many businesses run more than one agent: a front-desk agent and a sales line, or one agent per location. Cypher LM lets you decide where each file belongs.
Shared knowledge is a library every agent uses on every call. It is the right home for facts that are true across the business: company policies, opening hours, how to contact support. Update a shared file once and every agent uses the new version on its next call.
This agent's knowledge holds files only one agent uses. It is the right home for material that is specific to a role or a location: the sales agent's package details, or one branch's parking instructions. Keeping these files separate stops one agent from quoting another location's details.
Both libraries support folders, so you can organise files the way your team thinks about them: by service, by location or by department. Moving a folder never deletes its files; if you remove a folder, its contents move up a level. When you remove a single file, the agent simply stops using it on new calls, and you can upload it again at any time.
Importing from Google Docs, Sheets and Drive
If your team already keeps its price lists and policies in Google Workspace, you do not need to download and re-upload them. Connect Google in Cypher LM and use Import from Google in the agent's knowledge section.
You can paste a Google Docs link directly, which works with Google Docs access alone. To search your Drive and import Google Sheets, PDFs and other files, connect Google Drive as well. Imported Sheets come in as tabular text, which works well for price lists and service catalogues. The same 4 MB size limit applies.
The real advantage of importing is that the document stays where your team edits it. Each imported file shows its sync status, and you can press Resync after a change. Cypher LM also checks imported files regularly and re-imports the ones that changed, so the agent stays close to the version your team maintains. If you remove an imported file from the agent, the original stays untouched in your Google Drive.
Imported documents also help beyond the voice agent. Cypher M, the Cypher LM live call copilot, uses your business knowledge when it suggests what to say on live calls, so the same well-written FAQ that helps your agent also helps your human team.
How to test your knowledge base before customers do
Never let customers be the first to discover a gap. A short, deliberate test session catches most problems in under an hour.
- List twenty real questions. Use the same list your front desk wrote for the FAQ, plus five awkward ones: price exceptions, complaints, out-of-hours requests.
- Ask them in text first. The agent's Knowledge pages include Test a question, a text preview of how the agent would answer, which is a fast way to check facts without placing a call.
- Then make test calls. Use Test call or the agent's own page and ask the same questions out loud, including the way a hurried caller would phrase them.
- Read the transcripts. Every test call appears in Calls with a transcript, so you can see exactly what was said.
- Fix the document, not the symptom. If an answer was wrong, find the passage it came from and rewrite it. If an answer was missing, add it to the right file.
- Repeat the questions that failed. A change only counts once the agent gets the question right.
Pay attention to answers that are technically correct but unhelpful. If a caller asks about parking and the agent reads the full address twice, the fix is a clearer parking section, not a different agent.
Keeping it fresh after launch
A knowledge base is never finished. Prices change, staff change, and customers find new questions to ask. The best source of improvements is your own call history.
Read a handful of call recaps every week. Each recap in Calls › Recaps gives you the summary, key points, action items and a follow-up draft. When you see the same open question appear across several calls, or an action item that exists only because the agent could not answer, that is a gap in the knowledge base. Add the answer and the follow-ups disappear.
Set a simple routine. Whenever a price or policy changes, update the source document the same day; if it is imported from Google, resync it, and if it was uploaded, upload the new version and remove the old one. Once a quarter, review every file for contradictions and remove anything outdated.
Give one person ownership. Knowledge bases decay when everyone assumes someone else is updating them. A named owner, even for a few minutes a week, keeps the agent accurate.
How this works in Cypher LM
Here is the full path from an empty agent to one that answers from your own documents.
- Build the agent. The setup wizard asks for your business website, details, audience, language and the agent's identity, and reads your website to get started.
- Add knowledge. Open the agent, go to Knowledge, and upload PDF, DOCX, CSV or TXT files (up to 4 MB each, 10 at a time), or use Import from Google for Docs, Sheets and PDFs.
- Choose the scope. Put company-wide files in Shared knowledge and role- or location-specific files in this agent's knowledge, using folders to stay organised.
- Preview and test. Check the text preview of each file, try questions in the text preview, and run test calls; review the transcripts in Calls.
- Go live. Deploy the agent on your website with the widget, and keep improving the knowledge base from what recaps show.
Files you add are used on the agent's next call; there is no separate publishing step. If the agent is not live yet, it starts using the files as soon as it is.
Frequently asked questions
How many files should my knowledge base have?
There is no target number. Most businesses start with four to six focused documents (services and prices, hours, policies, FAQs) and add more as call recaps show gaps. Fewer, clearer files work better than many overlapping ones.
Does the agent read images or scanned pages?
The agent reads the text in a file. If a PDF is a scan or a picture of a menu, the text may not come through. Use the file preview to check, and if it is empty, retype the content or export it as a text-based PDF.
Can I import a Google Sheet as a price list?
Yes. Connect Google Drive, choose the Sheet in Import from Google, and it is imported as tabular text. Use clear column headings such as Service, Duration and Price so each row reads as a complete answer.
What happens when I update a Google Doc?
Imported files show their sync status. You can resync a file yourself after an edit, and Cypher LM also re-imports files that have changed. The agent uses the updated version on new calls.
Will the agent make up an answer if it is not in the documents?
You decide what the agent does when it does not know. Set a clear fallback in the agent's behaviour settings, such as offering to have the team follow up, and the call recap will show the open question so you can add the answer.
Conclusion: treat your knowledge base like a product
The businesses that get the most from a voice agent treat the knowledge base as a living product: short, specific documents, one answer per question, no contradictions, and a weekly look at what callers asked that the agent could not answer.
None of this requires technical skill. It requires the same care you would give to training a new member of staff. Start with your FAQ, your prices and your policies, test with real questions, and let your call recaps tell you what to write next.
If you want to see how your own documents perform, start a free trial, build an agent from your website and upload your FAQ. Or book a demo and we will walk through a knowledge base with you.
