What an agent reads, writes and protects while a conversation runs, one section per term.

The knowledge base

Knowledge base

The retrieval corpus. Documents the agent searches to ground an answer. Two words, never one. Every environment shares one knowledge base but decides which documents it includes.

Also asked as:

the documents the assistant answers from
upload our help articles so the bot can use them
ground replies in our own content

Related: Data source · Document · Knowledge base tool · RAG

Documented in: Querying the knowledge base · Importing data sources › Knowledge base and environments

Data source

Where knowledge base content comes from: a file, a URL, or a synced integration. Web pages, sitemaps, uploaded files and integrations such as Zendesk Help Center or Shopify products are all data sources.

Also asked as:

import a website into the knowledge base
add PDFs for the assistant to read
crawl our help center

Related: Knowledge base · Refresh rate · Metadata

Documented in: Importing data sources › Adding a data source · Importing data sources › Data types · Importing data sources › Troubleshooting imports · Shopify tool › Importing Shopify products into your Knowledge Base · Zendesk tool › Syncing Zendesk Help Center with your Knowledge Base

Document

One source inside the knowledge base: a file, a page, or a synced record. Retrieval returns passages from documents, not whole documents.

Also asked as:

one item in the knowledge base
the unit the bot searches over
a single file or page in the knowledge base
delete or replace one uploaded item

Related: Data source · Chunk

Documented in: Importing data sources › Data types · Importing data sources › Developers

Refresh rate

How often a URL or integration data source is re-synced into the knowledge base, set on import or afterwards.

Also asked as:

keep the imported pages in sync with our site
re-crawl a URL on a schedule
the knowledge base copy of our site is out of date
our crawled pages are stale after the site changed
re-sync pages pulled in from a URL
content we crawled has changed on the source site

Related: Data source

Documented in: Importing data sources › Refresh rate

Chunking strategy

How a document is split into the passages retrieval returns. LLM chunking strategies use a model to split content along its meaning rather than at fixed lengths.

Also asked as:

the bot returns half a paragraph
split documents more sensibly for search
passages cut in the wrong place

Related: Chunk · Knowledge base tool

Documented in: Importing data sources › LLM chunking strategies

Metadata

Tags attached to a data source so the knowledge base tool can filter what is returned — by brand, product line, locale or support tier.

Also asked as:

only search the documents for one product line
tag content by region or language
filter which sources a question is answered from

Related: Data source · Knowledge base tool

Documented in: Importing data sources › Metadata · Querying the knowledge base › Meta data filtering

Custom query

Overriding the knowledge base tool’s default search text (the user’s last message) with a specific variable or phrase; query re-writing lets the model rephrase the message before searching.

Also asked as:

search the knowledge base on something other than what the user typed
rewrite vague questions before retrieval
the bot searches with the wrong phrasing

Related: Knowledge base tool

Documented in: Querying the knowledge base › Custom query · Querying the knowledge base › Query re-writing

Chunk limit

How many content chunks the knowledge base tool returns per query (1–10, default 3). More chunks give more context but add latency and token usage.

Also asked as:

return more passages per question
the assistant only sees part of the answer
how many snippets the bot reads per question
give the assistant more context per query

Related: Chunk · Knowledge base tool

Documented in: Querying the knowledge base › Chunk limit

Source URLs

A chat-project option that makes the agent include the source URL alongside its reply so users can verify or read more. Tool messages, similarly, script what the user sees while the search runs.

Also asked as:

show a link to where the answer came from
cite the page behind a reply
a “searching…” message while the bot looks things up

Related: Knowledge base tool

Documented in: Querying the knowledge base › Show source URL(s) · Querying the knowledge base › Tool messages

Data

Variable

A named slot of project state an agent reads and writes during a conversation. Inserted anywhere by typing {, and set by the Set step, the Code step, tool responses or the widget.

Also asked as:

store what the customer told us for later
pass a value between steps
a field the assistant fills in during the chat

Related: Built-in variables · Persistent variable · Secret · Set step

Documented in: Variables › Creating variables · Variables › Using variables · Variables › Setting variable values

Persistent variable

A variable whose value carries over across sessions for the same user, chosen per variable when it is created or edited.

Also asked as:

remember a detail about a customer between visits
keep a value from one chat to the next without asking again
data that survives the end of a session
remember something about a customer across separate chats
don’t ask for the same detail again next time

Related: Variable · User ID · Chat persistence

Documented in: Variables › Persisting variables across sessions

Built-in variables

Variables every project has, set automatically when a conversation begins or an event occurs: user_id, vf_memory, vf_now, last_event and others.

Also asked as:

the system variables I can read
get the current date inside the agent
which values Voiceflow fills in for me

Related: Variable · User ID · Timezone

Documented in: Variables › Built-in variables

Secret

A credential stored for the agent to use without exposing it in a prompt or a function. Created on the Secrets page with a name, a value and a visibility (masked or restricted), and inserted from the Secrets tab of the { menu. The name is shared across the project; each environment holds its own value.

Also asked as:

store an API key securely for the assistant to use
keep a password out of the prompt
reference a credential in a request header

Related: Secrets across environments · API tool · Variable

Documented in: Secrets › Creating secrets · Secrets › Using secrets · Secrets › Security

Secrets across environments

A secret’s name is shared across the project, but each environment holds its own values: a default value its published version uses, and an optional draft value its draft version uses instead. Nothing fills in a value an environment doesn’t have; a missing value is sent as the placeholder text. A merge leaves Main’s values alone unless you choose to override them; duplicating a project copies Main’s values within the workspace; exporting copies names only.

Also asked as:

one credential name with a different value for staging and production
use a sandbox key while testing without changing the live key
why does my secret say missing value in another environment
the secret values disappeared after I exported the project
a credential that differs per environment under one name

Related: Secret · Environment · Merging

Documented in: Secrets › Secrets in each environment · Secrets › Testing with a non-production credential · Secrets › Merging to Main · Secrets › Duplicating and exporting projects

Persona

A saved set of traits used to simulate a particular kind of user when testing. Personas are scoped to the environment they were created in.

Also asked as:

test the assistant as a specific kind of customer
role-play a difficult caller when trying the agent
pretend to be a certain type of user
act like an impatient customer when testing the agent
simulate a specific type of user in a test
try the bot as if I were a difficult caller

Related: Test · Environment

Documented in: Personas · Personas › Personas and environments

Privacy

PII redaction

Stripping personal data from what gets stored. A model hosted inside Voiceflow’s infrastructure detects personal information in both the agent’s responses and the user’s inputs; production conversations are redacted, development ones are not.

Also asked as:

mask phone numbers and emails in transcripts
are customer details stored in plain text
strip personal data before it is saved
are phone numbers and emails stored in plain text in transcripts
hide personal details from the conversation logs

Documented in: PII redaction › How it works · PII redaction › Enabling PII redaction · PII redaction › Limitations · PII redaction › Viewing unredacted transcripts

Detected PII types

The more than 35 kinds of personal information redaction detects: identity (names, date of birth, age), contact details, financial and account numbers, and more.

Also asked as:

which sensitive fields get masked
does redaction catch credit card numbers
what counts as personal data

Related: PII redaction · Protection methods

Documented in: PII redaction › Detected PII types · PII redaction › What gets redacted

Protection methods

The four ways redaction handles content: censor (replace with a marker like [NAME]), erase, override and rich-text censoring, chosen by the type of content.

Also asked as:

what replaces the redacted text
markers like NAME or EMAIL in transcripts
how redacted names and numbers appear in transcripts
censor versus erase in redaction

Related: PII redaction · Detected PII types

Documented in: PII redaction › Protection methods · PII redaction › Redaction examples