A voice agent is only as good as what you give it to read. Everything your agent says on a call comes from one of three places: the business details and instructions you set when you built it, the files you upload to its knowledge base, and the question-and-answer entries you add yourself. Nothing else. Your agent cannot look things up on the internet mid-call, and cannot guess at a price or a policy it was never told.
This guide is about the middle two: the files and the FAQ. It covers what your agent can and cannot realistically know, how to audit the documents you already have before you upload them, how to write an FAQ document in a format an agent answers well from, how to structure price lists and catalogues as CSV files, how the website field and Google Drive import work, how to organise files into folders, and how to test your agent's answers with a question bank before customers do it for you.
None of this requires technical skill. It requires the same discipline as training a new front-desk hire: write things down clearly, keep them current, and check the answers before you trust them. By the end of this guide you will have a repeatable process for building and maintaining a knowledge base that holds up call after call.
01
What your agent can and cannot know
Every answer your agent gives on a call is assembled from three layers, in roughly this order of weight: the instructions and business details you set up when you built the agent (its name, tone, what it does and does not handle), the knowledge files you upload on its Knowledge pages, and the FAQ entries you add by hand. There is no fourth source. Your agent does not browse the web during a call, does not check your point-of-sale system, and does not remember anything from a caller's previous conversation unless that detail was written into one of the three layers first.
The three layers, and what each is good for
Business details and instructions are best for things that rarely change and describe the business as a whole: your name, what you sell, your tone, your escalation rules. Knowledge files are best for anything long or likely to be referenced more than once: a full price list, a policy document, a catalogue, a staff directory. FAQ entries are best for short, specific questions and answers callers actually ask, that a longer document answers poorly or not at all.
Shared library versus a single agent's own files
From an agent's Knowledge tab, files can be added to two places. "Shared knowledge" sits in one library every agent on the account can draw from, useful for policies, pricing or background that applies across your whole business. A file added directly to one agent instead stays private to it, useful when a second agent serves a different location or product line with different answers. When you add a file you are asked who should use it; choose shared only for material genuinely true for every agent you run.
What your agent genuinely cannot know
An agent cannot know anything that was not written into its instructions, its files, or its FAQ. In practice this means three common gaps worth planning around. First, live or changing data: stock on hand, whether a specific appointment slot is free, whether a particular order has shipped. Unless that information comes from a connected tool such as your calendar, treat your agent's knowledge as a snapshot, not a live feed, and write it that way ("most orders ship within two business days" rather than a number that changes weekly). Second, anything outside what you uploaded: an agent will not invent a return policy if you never described one, but it may try to be helpful with a vague, generic answer, which is why a clear fallback instruction matters as much as the files themselves. Third, anything said on a different call: each conversation starts fresh; a caller who mentioned their order number yesterday has to mention it again today unless that detail was saved elsewhere in your business systems.
Why format matters more than volume
A ninety-page policy manual dropped in as one PDF is harder for an agent to answer from precisely than the same content broken into short, clearly headed sections, because the agent has to find the right passage among a lot of surrounding text. Shorter, well-labelled documents, plus a dedicated FAQ for questions that come up constantly, consistently produce more accurate answers than one long file trying to cover everything. The rest of this guide is about getting that structure right before you upload anything.
A 12-chair dental clinic
A 12-chair dental clinic uploads its entire new-patient handbook as a single 40-page PDF, expecting the agent to answer anything a patient might ask. In practice, callers ask the same six or seven questions: hours, whether they take a specific insurance, whether X-rays are included, cancellation notice, and parking. The clinic later adds those seven as FAQ entries and trims the handbook PDF down to the sections patients are actually calling about, and the agent's answers become noticeably sharper.
Before you upload anything, decide
- Is this fact permanent enough to live in the agent's instructions, or does it belong in a file?
- Should this file be shared across every agent, or does it belong to one agent only?
- Is this long-form reference material (a file) or a short, specific Q&A (an FAQ entry)?
- Does any part of this document describe something that changes weekly or daily? If so, note that it may go stale.
- Have you written a clear fallback line for anything the agent is not given an answer to?
Getting this split right, instructions for the stable facts, files for the detailed reference material, FAQ for the specific recurring questions, is the single biggest factor in how reliable your agent sounds. The next section covers how to audit what you already have before any of it goes near the upload button.
02
Auditing your source material before you upload
Most businesses already have the raw material for a strong knowledge base sitting somewhere: a website, an old FAQ page, a price sheet in a spreadsheet, a staff handbook, a policy email that gets forwarded to new hires. The work is rarely about writing from scratch; it is about pulling that material together, removing what is out of date or contradictory, and reshaping it into a form an agent can answer from cleanly. Do this audit before you touch the upload button, not after.
Start by listing every source you already have
Go wider than you expect to need. Typical sources include your website's About, Services, Pricing and FAQ pages, any printed or PDF price list, a cancellation or refund policy, a product or service catalogue, onboarding material for new staff, and a running list of the questions your front desk or support inbox answers most often. That last one is frequently the most valuable and most overlooked: whoever answers your phones today already knows which five questions eat most of their time.
Resolve conflicts before you upload, not after
The most common cause of an agent giving a wrong answer is not a bad agent, it is two documents that disagree. A price list from eighteen months ago says one number; your current rate card says another; both end up uploaded, and the agent has to guess which one is current (or worse, blends them). Before uploading, decide which source is authoritative for each topic and retire or correct the others. If a document is partly outdated, fix the outdated parts rather than uploading it as-is and hoping the agent picks the right passage.
Check for currency, not just accuracy
A document can be accurate and still misleading if it does not say when it was last true. "Delivery takes 3 to 5 business days" is fine; the same line left unedited for two years after your delivery process changed is a liability. As you audit, flag anything with a date-sensitive fact, a price, a staff name, an hours note, so you know what needs rechecking on a schedule, not just once.
Write for the agent, not for a human skimming a PDF
A human reader fills in gaps using context and common sense; an agent answers more literally from what is on the page. A policy written as flowing prose ("We generally try to accommodate most reasonable requests where possible, subject to availability") reads fine to a person and gives an agent almost nothing concrete to repeat. The same policy rewritten directly ("Cancellations made more than 24 hours before the appointment are free. Cancellations within 24 hours are charged 50% of the service fee.") gives the agent an answer it can state with confidence.
Group by topic before you split into files
Once conflicts are resolved and language is tightened, group the material by topic: hours and location, pricing, policies, products or services, and anything specific to a single agent rather than the whole business. This grouping maps directly onto how you will structure files and folders later, and is much easier to do on a working document than after everything is already uploaded.
A two-person recruiting agency
A two-person recruiting agency finds that its "services" page, an old client FAQ from two years ago, and a fee schedule in a spreadsheet all describe placement fees slightly differently, because the fee schedule was updated last but the other two documents never were. Before uploading anything, they retire the old FAQ, update the services page to match the current fee schedule, and keep the spreadsheet as the single source of truth for numbers, converting it to a CSV for upload.
Knowledge audit checklist
- List every existing source: website pages, price lists, policies, catalogues, staff handbooks, support inbox patterns.
- Identify the single authoritative source for each topic (pricing, hours, policies) and note which others to retire.
- Flag anything with a date-sensitive fact (price, hours, named staff, delivery time) for a recheck schedule.
- Rewrite vague or conditional language as direct, literal statements an agent can repeat.
- Group remaining material by topic: hours and location, pricing, policies, products and services, agent-specific detail.
- Note which grouped topics apply to every agent (shared) and which apply to one agent only.
A clean audit takes an afternoon for a small business and pays for itself the first week the agent is live, because almost every "wrong answer" a customer reports traces back to source material that was out of date, contradictory, or too vague to act on. The next section turns this audited material into an FAQ document the agent can answer from precisely.
03
Writing an FAQ document your agent answers well
The single highest-value addition to a knowledge base is a direct FAQ: real questions, each followed by a specific, complete answer. Agents answer FAQ-style content more reliably than long prose for a simple reason: there is no passage to search and interpret, the exact question is already matched to the exact answer you want repeated.
Format: one question, one direct answer
Write each entry as a question a real caller would ask, in the words they would actually use, followed by the answer you want given, in full sentences, as if the agent is speaking it aloud. Avoid answers that only make sense next to a table or a diagram; an agent on a voice call can only say words. Keep each answer to two or three sentences. If an answer genuinely needs more detail, that topic likely belongs in a longer knowledge file rather than the FAQ.
Use the caller's words, not your internal terms
Customers do not ask about your "service tiers" or "SKU catalogue," they ask "how much does it cost" and "do you have that in stock." Write questions the way customers phrase them, including informal versions ("what time do you close" as well as "what are your hours"), because the closer your written question is to what gets asked on a real call, the better the match.
Group related questions together
A long flat list of forty questions is harder to maintain than the same forty questions grouped under short headings: Hours and Location, Pricing, Appointments and Cancellations, Products and Services, Policies. Groupings also make gaps easier to spot, if "Pricing" only has one question but your inbox gets pricing questions daily, more entries are needed there.
Keep answers current and specific
An answer like "pricing varies" is technically true and practically useless; a caller asking about price wants a number or a range. Where a true price or policy exists, state it. Where it genuinely varies, say what it depends on ("the fee depends on the size of the property; a two-bedroom apartment is around X") rather than leaving the caller with nothing. Review FAQ answers on the same schedule you review prices and policies elsewhere.
Use your agent's own tools to find gaps
Two tools on an agent's Knowledge pages find FAQ gaps after launch. "Couldn't answer" records real questions the agent could not answer, with how many times and when each was last asked; turning the most frequent into FAQ entries closes real gaps rather than guessed ones. "Test a question" lets you type a question and see the answer the agent gives right now, useful for checking a new FAQ entry gets matched before a real caller finds out it doesn't.
A template to start from
The structure below works for almost any small or mid-sized business. Copy it, replace the bracketed placeholders, and add or remove groups to match your own services.
Template FAQ document
FAQ document checklist
- Every question is phrased the way a real customer would ask it, informally included.
- Every answer is two to three sentences, written to be spoken aloud.
- No answer says "it depends" without also saying what it depends on.
- Related questions are grouped under short topic headings.
- Prices, hours and policy numbers match your current, audited source documents.
- The ten most common questions from your phone or support inbox are all covered.
A single-location boutique gym
A single-location boutique gym starts with eight FAQ entries covering membership pricing, class schedules and the cancellation window for personal training sessions. After two weeks, its "Couldn't answer" list shows twenty-two callers asked about guest passes, a question the gym had not thought to include. Adding one FAQ entry for guest passes closes the gap the same day, rather than waiting for a pattern of complaints.
An FAQ built this way, specific questions in the caller's own words, direct answers written to be spoken, grouped and kept current, does more to improve accuracy than almost any other single change you can make. The next section covers the other half of most knowledge bases: price lists, policies and catalogues as files rather than FAQ entries.
Put this guide to work
Build it in your own account
Start a 7-day free trial with every Cypher LM product included, or book a demo and we will walk you through it on a live call.
04
Structuring price lists, policies and catalogues
Not everything belongs in the FAQ. Price lists with many line items, catalogues, and longer policy documents are better uploaded as files, since they are reference material an agent searches rather than a short list of fixed questions. The format you choose has a real effect on how accurately the agent reads it back.
Price lists and catalogues: use CSV
A spreadsheet or CSV file, one row per item, with clearly named columns, is the most reliable format for anything list-like: services and prices, products and prices, a menu, a rate card. Keep columns short and literal. A CSV with columns like Service, Price, Duration, Notes is far easier for an agent to answer from than the same information as flowing paragraphs, because each row is a self-contained fact.
CSV price list template
CSV catalogue template
A CSV with an "In stock" or availability column is worth a note: unless that file is kept current, write the column as "call to confirm" rather than a true/false flag that will silently go stale. An agent repeating a stale "yes, in stock" with total confidence is worse than one that honestly says stock levels change and offers to have someone confirm.
Policies: short sections beat long prose
A refund, cancellation, warranty or terms-of-service policy uploaded as a DOCX or PDF works best broken into short, clearly headed sections, one policy per heading, written as direct statements rather than legal-style conditional prose. A document written for a lawyer to review and one written for an agent to answer from aloud are different documents; keep the legal version for your records and write a second, plainer version for the knowledge base.
Catalogues with many items: split by category
A catalogue of dozens or hundreds of items is easier to manage as several smaller files organised into folders by category (services, products, add-ons) rather than one enormous file. This also makes updates safer: changing this month's seasonal menu only touches one file, not the entire catalogue.
Matching source type to format
The table below is a quick reference for which format to use for different kinds of source material.
| Source type | Best format | Where it lives |
|---|---|---|
| Price list, rate card | CSV, one row per item | Knowledge file (shared or agent) |
| Product or service catalogue | CSV, split by category if large | Knowledge file, organised in folders |
| Policy (refund, cancellation, terms) | DOCX or PDF, short headed sections | Knowledge file (usually shared) |
| Recurring, specific customer questions | Direct Q/A pairs | FAQ entries |
| Business description, tone, escalation rules | Plain instructions | Agent instructions, set in the wizard |
| Hours, address, main phone number | Plain instructions or FAQ | Agent instructions and FAQ |
File size and type limits
Uploaded files must be PDF, DOCX, CSV or TXT, under 4 MB each, and you can add up to 10 files in a single upload. Larger documents should be split rather than compressed; a 4 MB limit is generous for a text-based price list or policy document, and splitting a long catalogue by category usually keeps every file under that limit while also making each one easier to update independently.
Before uploading a price list, catalogue or policy
- Line-item data (prices, products, menu items) is in CSV, one row per item, with short, literal column names.
- Policies are rewritten as short, direct statements grouped under clear headings, not legal-style prose.
- Large catalogues are split by category into smaller files rather than one very large file.
- Anything genuinely time-sensitive (stock, seasonal pricing) is marked "call to confirm" rather than stated as fixed fact.
- Every file is under 4 MB and in PDF, DOCX, CSV or TXT format.
A mid-sized furniture retailer
A mid-sized furniture retailer initially uploads its entire 90-item catalogue as one PDF exported from its website. Several callers are told incorrect dimensions for items the agent had to search for inside a long, inconsistently formatted document. Converting the catalogue to a CSV with columns for Item, Category, Price, Dimensions and In stock, split into three files by room type, resolves the dimension errors because each fact now sits in its own clearly labelled cell rather than inside a paragraph.
Formatting price lists and catalogues as structured CSV data, and policies as short direct sections, is the difference between an agent that answers confidently and one that paraphrases uncertainly. The next section covers two other sources worth knowing about: your website and Google Drive, plus how folders keep a larger knowledge base organised.
05
Website, Google Drive import, and organising with folders
Two sources fill out a knowledge base beyond manually uploaded files: the business website entered when you first built your agent, and a direct import from Google Drive. Used well, both save real time; used without review, both can quietly carry forward the same out-of-date or contradictory content this guide has already warned against.
The website field: a starting point, not a finished answer
When you build an agent, entering your business website lets the wizard visit the site and pre-fill several business-detail answers automatically, saving you from retyping content that already exists on your public pages. This is a genuine time-saver for the first draft, but it is a starting point, not a finished answer: review every pre-filled field before moving on, the way you would review any first draft. Website copy is frequently written for marketing rather than for a caller asking a direct question, and it can be out of date in the same ways any other source material can be.
Importing from Google Drive
From an agent's Knowledge pages, "Import from Google Drive" lets you search your connected Drive and bring a Google Doc, Sheet or PDF directly into the knowledge base, or paste a direct link instead of searching. A Sheet imports as structured data, similar to a CSV, which makes Drive import a convenient way to keep a live price list in Sheets and bring it in without exporting a file by hand each time. Google must be connected first; a connection showing as needing reconnection must be re-authorised before an import will succeed.
Keeping an imported file current with resync
A file imported from Drive stays linked to its source. Each imported file's menu offers "Resync from Google," which re-checks the original file and pulls in any changes since the last import, and "Open," which jumps to the source file in Drive. This is the practical advantage of Drive import over a manual upload: when your team edits the live price list Sheet, resyncing brings that edit in without re-exporting and re-uploading a CSV by hand. Changes in Drive do not reach the agent until you, or the scheduled recheck, resync.
Organising with folders
As a knowledge base grows past a handful of files, folders keep it navigable. On an agent's own Knowledge page you can create folders, rename them, move files or whole subfolders between them, and delete folders you no longer need. A sensible starting structure mirrors the grouping suggested earlier in this guide: Pricing, Policies, Products or Services, and anything specific to one location or team. Folders also make it immediately obvious, at a glance, whether a topic is thin (an empty or near-empty Policies folder) before a caller discovers the gap first.
Setting up folders for a growing knowledge base
- Create one folder per major topic (Pricing, Policies, Products/Services, Location-specific).
- Move existing files into the matching folder rather than leaving everything at the top level.
- Check each folder has at least one file; an empty folder is a visible reminder of a gap to fill.
- For multi-agent accounts, decide per folder whether its contents belong in the shared library or a single agent's own files.
- Revisit the folder structure whenever you add a new product line, service or location.
Shared library versus per-agent files, in practice
If you run more than one agent, for example one per location or product line, the shared knowledge library is the right place for anything true everywhere: company-wide policies, a brand-wide FAQ, core pricing that does not vary by location. Each agent's own files are the right place for anything that varies: a location's address and hours, a product line specific to one agent, local promotions. Reviewing this split after adding a new agent avoids duplicating shared content separately into every agent's own files, which then drifts out of sync the first time one copy is updated and the others are not.
A property management company with three agents
A property management company runs three agents, one per building, sharing the same company-wide maintenance request policy and tenant FAQ, but each with its own building-specific contact number and amenity list. Keeping the policy and FAQ in the shared library, and only the contact number and amenities in each agent's own files, means a change to the maintenance policy only has to be made once and instantly applies to all three buildings' agents.
Between website auto-fill for a fast first draft, Drive import for material that changes in place, manual upload for everything else, and folders to keep it all navigable, most businesses can build a complete, well-organised knowledge base within an afternoon. The final section covers how to confirm it actually works before customers are the ones testing it.
06
Testing your agent's answers and fixing what's wrong
A knowledge base that looks complete and a knowledge base that actually answers correctly are two different things, and the only way to tell them apart is to test it, deliberately and before real callers do. This closing section covers how to build a test-question bank, how to use the built-in test tool, how to read the "Couldn't answer" log, and how to fix a wrong answer once you find one.
Build a test-question bank before you go live
A test-question bank is a written list of questions you expect callers to ask, covering every topic your knowledge base handles, worked through deliberately rather than discovered live. Pull questions from three places: the FAQ you wrote, the files you uploaded (pick specific facts from each price list or policy and phrase them as questions), and the common-questions list from whoever currently answers your phones. A good bank covers hours and location, pricing, your most-asked policy questions, and at least one question for every file you uploaded.
A bilingual-ready test-question bank
- One question for each FAQ group (hours, pricing, appointments, policies).
- At least one question pulled directly from every uploaded price list or catalogue.
- At least one question pulled directly from every policy document.
- A deliberately vague or off-topic question, to check the fallback response is used correctly.
- A question phrased informally, the way a real caller would actually say it, not textbook-perfect.
- A question about something genuinely not in the knowledge base, to confirm the agent admits it rather than guessing.
Using the built-in "Test a question" tool
Each agent's Knowledge pages include a "Test a question" tool: type a question, and it shows the answer the agent would give right now, using its current instructions, files and FAQ, without needing to place an actual call. When the answer matches an existing FAQ entry, the tool shows which one matched, with a shortcut straight into editing it. Work through your entire test-question bank this way before considering the knowledge base finished, it takes a few minutes and catches most problems before a customer does.
Reading the "Couldn't answer" log after launch
Testing before launch catches obvious gaps; the "Couldn't answer" log catches the ones you did not think to test. Every question the agent could not answer is recorded here, with how many times and when it was last asked, so you can prioritise by actual call volume rather than guesswork. Reviewing this list weekly in the first month, then monthly afterward, is the single best habit for keeping accuracy high as real call patterns reveal gaps your original audit missed.
Fixing a wrong answer, step by step
When a test, or a real call, surfaces a wrong or unclear answer, work through it in order: first check whether the underlying fact is correct in your source material (the agent may be repeating an outdated price accurately); second, if the fact is correct but phrased poorly, add or rewrite an FAQ entry with a clearer, direct answer, since specific FAQ entries generally beat a longer document's ambiguity; third, if the agent could not find the answer at all, check whether the file was uploaded to the right scope (shared versus this agent only) and folder; fourth, re-test with the "Test a question" tool to confirm the fix worked.
A small accounting firm
A small accounting firm tests its agent before launch and finds that a question phrased as "do you do taxes for freelancers" returns a vague answer, even though the firm's services file clearly lists freelance tax preparation. Rather than rewriting the file, the firm adds a direct FAQ entry using that exact phrasing, since the gap was in matching the caller's words, not in the underlying information. Retesting the same question afterward confirms the fix.
What to do next
Building and maintaining a knowledge base is not a one-time task, it is a short routine worth repeating. To put this guide into practice:
- Run the knowledge audit from section two on everything you already have, and resolve any conflicts between sources before uploading anything.
- Write an FAQ document using the template in section three, covering at minimum your ten most common questions, grouped by topic.
- Convert any price list or catalogue to CSV, and rewrite policies as short, direct sections, following the formats in section four.
- Decide, file by file, what belongs in the shared library versus a single agent's own files, and organise everything into folders.
- Build a test-question bank and work through it with the "Test a question" tool before you consider the agent ready for real callers.
- Review the "Couldn't answer" log weekly for the first month, turning the most frequent real gaps into new FAQ entries.
A knowledge base built this way, audited sources, a direct FAQ, structured price lists, sensible folders, a habit of testing and reviewing, is what separates an agent that sounds confident because it knows the answer from one that is guessing. For the next step, see building your first AI voice agent and deploying your agent. If your callers speak Urdu as well as English, the guide to bilingual agents covers preparing knowledge material in both languages.
