Handbook
Genesys Cloud knowledge: the workbench and knowledge fabric handbook
A detailed handbook for the people who own knowledge in Genesys Cloud. It compares the knowledge workbench and knowledge fabric, then covers matching and confidence thresholds, content search and phrases, variations, connectors, chunking, filtering, where knowledge surfaces, bot flows, the optimizer, licensing, languages, permissions and governance. Checked against Genesys sources on 15 November 2026.
In short
- Choose the model first: workbench articles for bot flows and the portal, the fabric for generative answers from documents.
- Confidence ranks answers, it does not prove them; set thresholds by content type and language, using Genesys's defaults.
- Content search is chosen when a base is created, and inverts phrase practice: remove phrasings when it is on.
- Fabric content is cut into fixed 400-token chunks of text only, so write self-contained sections and describe tables in words.
- Name a content owner, knowledge manager and publisher, run the optimizer weekly, and keep your own record of AI answers.
01
What knowledge services are for, and the two platforms
This handbook is for the people who own knowledge in Genesys Cloud: knowledge managers, content owners, solution designers and the teams building bots and agent assistance. It covers the knowledge workbench, the knowledge fabric, how answers are matched and generated, where knowledge appears, how it is measured, and the governance that keeps it right. Every statement about Genesys Cloud comes from Genesys's documentation, listed in the sources and checked on 15 November 2026. Where something is our own practice, it is labelled QVCCS practice.
1.1 Three jobs, usually one person
Genesys describes the knowledge workbench as letting a knowledge author create and manage articles, fine-tune the search process and view the performance of knowledge bases. That sentence names three jobs: writing content, tuning search and watching performance. Most programmes fill only the first, which is why part 14 exists.
The same content now serves four consumers with different quality bars: customers in self-service, bots choosing an answer by confidence score, agents reading suggestions mid-conversation, and AI that rewrites retrieved content into a cited answer. The fourth changes the content problem. A badly written article used to produce a bad search result; now it can produce a fluent, cited and wrong answer.
1.2 Two platforms with different content models
Genesys offers two knowledge platforms, and choosing between them is the first design decision. With the knowledge workbench V2, you own structured question-and-answer articles inside Genesys Cloud, with phrases, categories, labels, versions and touchpoint variations. With the knowledge fabric, you connect document sources; Genesys ingests the text, divides it into chunks, indexes it and generates answers with cited sources. Genesys removed knowledge workbench V1 on 29 May 2023; that affected only Dialog Engine bot flows, not digital bot flows or V2.
| Knowledge workbench V2 | Knowledge fabric | |
|---|---|---|
| Source | Authored question-and-answer articles, or structured articles from Salesforce, ServiceNow or the universal connector | Files: .txt, .md, .doc, .docx, .csv, .xls, .xlsx, .html and .pdf |
| Articles copied into Genesys Cloud | Yes | No |
| Touchpoints | Agent Copilot, Virtual Agent, AI Guides, bot and digital bot flows, knowledge portal and app, agentic virtual agent | Agent Copilot, Virtual Agent, AI Guides, agentic virtual agent |
| Context-aware search | No | Yes |
| Source references in answers | Yes, with limited support | Yes |
| Primary use | Standard responses not generated by AI, or AI-generated responses from articles | AI-generated responses from fabric sources |
1.3 Two rows that decide most designs
The touchpoint row is a hard constraint, with one nuance. Genesys's comparison states that the knowledge fabric is currently not supported in bot and digital bot flows, while its Architect guidance shows a fabric configuration linked to a bot flow with Virtual Agent enabled. In practice, classic bot flows use workbench knowledge bases, and the fabric reaches flows only through Virtual Agent. An organisation whose self-service runs on bot flows without Virtual Agent cannot use the fabric there, and that is easy to miss after an answer-generation demonstration.
The copied row is a governance decision in disguise. Article connectors create a copy of each article in a Genesys Cloud knowledge base, so two copies exist with different owners. The fabric leaves the source authoritative and re-ingests it. If a compliance team will not accept a second copy of a policy document, that settles the question.
Genesys suggests the fabric when knowledge is held in documents or on websites, when you want to search several sources at once, need context-aware experiences or AI-generated answers, or want a modern self-service or agent assist experience.
02
Knowledge workbench V2: content model and limits
The workbench has four objects, and their relationships matter more than they first appear.
2.1 Knowledge bases, articles, phrases and categories
A knowledge base is a library of question-and-answer articles about a product, service, topic or category. Genesys's planning advice is to group similar articles into a base whose name shows the relationship, such as Asia Pacific office FAQs or shipping and returns FAQs. An article is a question-and-answer pair: when a customer asks something, Genesys's knowledge AI looks for a similar question and returns its answer. Phrases are alternative wordings of the same question (part 4). Categories and labels organise and filter content (part 5).
Content can be rich text, with images and videos through URLs rather than uploads. One sentence in Genesys's overview belongs in front of whoever will own the content: contributors should be well versed in the subject matter. Genesys is describing subject-matter authors, not a team transcribing someone else's knowledge.
2.2 The published limits
| Limit | Value |
|---|---|
| Knowledge bases per organisation | 500 |
| Question-and-answer pairs per knowledge base | 15,000 |
| Categories per knowledge base | 1,000 |
| Data format | JSON |
| Phrases per article, content search off | Unlimited |
| Phrases per article, content search on | 100 |
| Minimum length of a title or phrase | Three characters |
| Articles per bulk edit | Up to 100 |
- Genesys's documentation does not state limits on labels per base, category depth or article length; confirm any that a design depends on.
- The 100-phrase cap applies only with content search, which cannot be switched on for an existing base (part 4), so the two decisions must be made together.
2.3 A knowledge base is a unit of configuration
Bases are selected per touchpoint: a bot flow's default knowledge base, a portal's base, an Agent Copilot configuration. With limits this generous, the question is never capacity but blast radius: how much changes when one base is re-imported.
03
How matching works, and confidence thresholds
Genesys knowledge search uses AI to compare the questions it detects in a conversation with the articles in a base, assigns each a confidence level as a percentage, and shows the answers with the highest ratings. Understanding what that number does and does not mean is the basis of good tuning.
3.1 The top result is not always right
Genesys is plain about it: results are always sorted by confidence, but the article with the highest confidence might not be the answer the customer needs, and testing and modifying questions and answers improves results. Confidence is a ranking, not a probability of being correct, even though a customer may read it as one. Authors need to test real customer phrasings, and bot designs need sensible behaviour at medium confidence rather than always speaking the top match (part 11).
Genesys also notes that a knowledge base automatically learns more phrases for existing questions; adding phrases manually speeds that up, and the original question and answer stays unchanged.
3.2 The published thresholds
The Knowledge confidence threshold setting in Agent Copilot comes with Genesys's own defaults and a warning: a threshold too low increases true positives but also false positives; one too high reduces false positives but also true positives.
| Knowledge base type | Recommended default |
|---|---|
| Knowledge bases without content search | 0.5 |
| English content search knowledge bases | 0.65 |
| Non-English content search knowledge bases | 0.7 |
- For content search bases in Filipino, Greek, Hebrew and Malay, values above 0.7 may improve quality; Genesys suggests testing higher values without recommending a different default.
- If you use AI-generated answers, set the threshold to 0.65 or below for all languages.
- If a knowledge article supersedes an intent-based rule, lower the natural language understanding confidence threshold, so the bot does not ignore the article.
- Use phrases sparingly and evenly across articles; for content search bases, remove all phrasings and keep only title and body.
3.3 The threshold follows the content type
Read together, the guidance makes one point: the right threshold depends on content type and language, not taste. A base that moved to content search but kept the 0.5 default under-filters from that day, and the symptom is the AI answering the wrong thing, not an obvious threshold problem.
04
Content search and phrases
Content search sits between title-and-phrase matching and full fabric retrieval, and it is decided once, when a base is created. Phrases are the training mechanism for bases without it. The two need opposite disciplines.
4.1 What content search does
Content search looks for relevant information in the article body as well as titles and phrases, which Genesys says suits longer articles and third-party articles imported through a connector such as Salesforce or ServiceNow. It broadens what can be found without adding phrases to each article.
Genesys is equally clear about when not to use it: for FAQ-style content with well-defined phrases, title and phrase search is often enough, and Agent Assist manual search users, who usually know which article they need, are better served by title and phrase search.
4.2 It cannot be switched on later
Genesys states that you cannot automatically enable content search for existing articles. The documented route is to export the articles, create a new knowledge base with content search enabled, and import them. That is a rebuild, not a setting: a new base to relink in every touchpoint, and a decision about what happens to the old base's feedback and history. Content search bases are also the ones that appear in a flow's Natural Language Understanding > Knowledge > Default Knowledge Base list, so the choice affects where a base can be selected in Architect.
4.3 Phrases are where contact centre knowledge differs
Genesys defines phrases as alternative ways of expressing the same intent, and its example pairs one answer about logging in with questions about account details and balances. Phrases are added under the question's Phrasings tab, each of at least three characters, with an optional Autocomplete setting that offers the phrase as a prediction as the customer types.
A policy owner writes the question as what is the cancellation policy. Customers ask whether they can get out of their contract, how to stop the direct debit, or say they have moved house. Only real customer wording retrieves well, and only real transcripts and search logs contain it.
4.4 Sparingly and evenly
Genesys's tuning advice, use phrases sparingly and evenly distributed across articles, rewards unpacking. Sparingly, because every phrase is a route by which an article can catch the wrong question. Evenly, because if three articles carry hundreds of phrases and ninety carry one, the well-fed articles win searches they should not. Automatic learning does not replace review: it learns from whatever traffic arrives, so early mistakes compound.
05
Categories, labels and touchpoint variations
Categories, labels and variations decide how content is organised, filtered and tailored. Two documented rules here are easy to discover too late.
5.1 One category, many labels
Categories group articles with similar content, for example home loans or insurance policies, and can be arranged in a hierarchy; each article links to one category only. Labels tag content and make filtering easier, and an article can carry several. Categories are what customers browse in the knowledge portal, so they must follow the customer's view of the world. Labels are for internal filtering and segmentation, such as audience, review date or legal approval.
Categories and labels are created at base level under Knowledge > Categories & Labels, using Nest category under to build the hierarchy. Genesys's quick start places this before any articles exist, which is the right order: changing a hierarchy after content and portal configuration exist means reworking both.
5.2 Variations: one article, several renderings
Rather than duplicate articles, Genesys provides variations, which let you keep one question-and-answer article and change the answer by touchpoint or segment. The touchpoints are Agent Assist, Dialog Engine bot flows, digital bot flows, the knowledge portal and the knowledge app for Messenger. Genesys's use cases: a variation for Agent Assist tailors information for an agent, while variations for bots, portal and Messenger tailor the customer's self-service journey. One article, one owner, one review is better than two articles that drift apart and then contradict each other on a recorded call.
5.3 Three mechanics, one counter-intuitive
- Naming differs by route: new articles get a single tab titled Default; existing articles without variations get a tab titled Variation 1.
- Priority is positional. If several variations match, the knowledge base returns the rightmost tab, and to raise a tab's priority you drag it to the left of the others.
- Segments are a separate axis based on customer and journey attributes, and require Predictive Engagement to be enabled.
06
Lifecycle, bulk editing and import
Articles have a real content lifecycle, and the tools for changing many articles at once have behaviours to understand before anyone runs them against a live base.
6.1 Draft, publish, revise and restore
Articles are drafted and published separately. An article's Revisions tab shows each version, when it was saved, who saved it, and the number of days before that version is permanently deleted. Versions expire, so revision history is not an indefinite audit record, and Genesys's documentation does not state the retention period; read it in your organisation.
You can preview an earlier version, including its question and answer, status, feedback, touchpoints and segments, without being able to edit it, then restore it if published edits were unwanted. An article can also be included in or excluded from knowledge surfacing, which is the safe tool for content that is correct but should not be used for now.
6.2 Bulk editing behaviours
Bulk editing covers categories, labels, publishing, restoring and deleting for up to 100 articles at once. Three documented behaviours matter.
- Adding labels in bulk does not remove existing labels. Bulk labelling adds; it does not replace.
- Bulk restore returns articles that were never published to an unpublished state, so a restore across a mixed selection can take content out of service.
- Bulk publish does not change articles that were already published.
6.3 Import, export and environments
Articles can be imported and exported as JSON, which is how content arrives at scale, moves between environments, and how the content search rebuild in part 4 is done. Import and export jobs have their own permissions, separate from editing articles, and should sit with a platform administrator rather than every author.
Knowledge content does not travel with configuration. Architect flows move as YAML and other configuration through CX as Code, but articles move as JSON import and export. A development-to-production approach must say explicitly how knowledge content is promoted, or testing will pass against content that does not exist in production. With no documented scheduled export, export is also the backup, so give it an owner and a frequency.
07
Article connectors: Salesforce, ServiceNow and universal
Workbench connectors connect to structured knowledge articles in external systems: Salesforce, ServiceNow and a universal connector for articles. Genesys states that when you connect this way, a copy of each article is created in a Genesys Cloud knowledge base. Built-in connectors are managed from Admin > Knowledge > Sources, with Third party Sources and Activity tabs, so authors do not have to recreate content.
7.1 Genesys's own warning
The most useful paragraph in Genesys's connector documentation is a caution. Connecting an existing knowledge tool means you do not recreate or import content by hand, but it does not guarantee good answers: connectors bring articles over unchanged, many were never written with answer generation in mind, and Genesys advises reviewing and adapting content to get the best quality from features such as Agent Copilot.
That answers the most common business case for knowledge: we already have thousands of articles, so this is a configuration job. It is a content job with a quick way of getting the material in.
7.2 Practical constraints
- For ServiceNow, only articles using the standard template are supported; custom templates will not import properly. Ask for the template list in discovery.
- You must add the Salesforce or ServiceNow integration before you can add its content as a source, so the CRM integration is a prerequisite.
- Salesforce has its own page layout setup steps for the connector, which puts the Salesforce administrator on the critical path.
08
The knowledge fabric and chunking
The knowledge fabric is the newer platform, and the one most demonstrations are built on, so it needs describing in both directions: what it does, and what it does not.
8.1 What it is
Genesys describes the fabric as turning scattered, static content into connected knowledge that improves with use, unifying sources, enriching them with AI and giving context-aware answers in self-service and agent-assisted channels. It connects to sources such as SharePoint, structures and indexes the selected documents, and divides them into smaller pieces. Its configurations define which sources are queried, how results are generated and when conversation context applies, and a preview widget lets administrators test search and answer results before anything reaches a live touchpoint. A built-in way to test before release is unusual; use it in every acceptance test.
8.2 What it is not
- Not for classic bot flows: the fabric reaches flows only through Virtual Agent (part 1).
- Not for images: Genesys states the fabric currently ingests text only, so diagrams, screenshots and flowcharts contribute nothing.
- Not a search index you tune: chunking is fixed, so content structure is the main lever you have.
- Not available in every language (part 13).
8.3 Fixed chunks of 400 tokens
Genesys states that all ingested content is divided into fixed-size chunks of 400 tokens, regardless of the document's structure, where a token is typically a word or part of a word, and that it uses fixed chunking for consistency and predictable performance. Chunks are indexed with embeddings and metadata, such as source document and URL. Results are ranked by confidence, the top three chunks are returned by default with their scores and sources, and AI generates the answer from those chunks and the query, in the touchpoint's language, citing each source.
A 40-page PDF is therefore not one document to the fabric; it is a couple of hundred independent fragments, of which three may be used. Genesys notes that because chunking is by token count, some chunks start or end mid-sentence.
8.4 How to write for chunking
| Genesys's guidance | Why it matters |
|---|---|
| Write in clear, self-contained sections that can stand alone | A chunk saying it must be done within 30 days, with no subject, cannot be retrieved or used safely |
| Use headings and subheadings | They help chunks capture logical breaks and improve relevance |
| Avoid overly long paragraphs | Long blocks are more likely to split mid-sentence |
| Put key points at the start of sections | Front-loaded facts survive a chunk boundary falling mid-section |
| Describe critical table or image content in the surrounding text | Only text is indexed, so that content is otherwise invisible |
- Genesys's troubleshooting advice: if results miss context or answers are incomplete, review source documents for structure and clarity.
- Search queries need at least three characters in alphabetic languages and two in logographic scripts such as Chinese and Japanese.
8.5 Review content before ingesting it
The chunking guidance is the part of a fabric deployment that changes what a content team writes, and it is cheapest to apply before ingestion.
09
Fabric sources, configurations and filtering
The fabric's objects are sources, which hold content, and knowledge configurations, which decide how touchpoints use it. Filtering decides which content each touchpoint sees, and has one constraint that can reshape a SharePoint estate.
9.1 Ways to bring content in
Genesys documents three routes: SharePoint folders, connected through a SharePoint integration where you select the folders to synchronise; fabric connectors from Genesys professional services, listed in AppFoundry as the Knowledge Fabric Connector; and the File Connector API for programmatic synchronisation. With fabric connectors, Genesys notes that file text is ingested directly. Treat any third-party connector as software needing the same security review as any integration.
9.2 Whoever can edit the folder publishes to customers
SharePoint is the usual source for UK organisations, and it brings a governance problem. A fabric source is a folder, so anyone who can write to it can change what the AI tells customers, including content never meant as knowledge: draft policy, meeting notes, an old price list, a document long past its review date. That is a different risk from curated articles, and part 14 treats it as the central control question.
9.3 Knowledge configurations
A knowledge configuration defines how knowledge is accessed and presented across touchpoints such as Agent Copilot and Virtual Agent, and it is what touchpoints reference. Two prerequisites trip people up: only sources that are indexed and ready to search appear when building a configuration, so a source still ingesting looks like a fault; and Genesys recommends matching the knowledge language to the touchpoint language, saying an English source with a non-English Virtual Agent is technically possible but not officially supported.
The configurations list shows each configuration's linked touchpoints, which is the impact check before any edit. Genesys recommends previewing a configuration before selecting it in a touchpoint, and configurations also supply AI Guides with data sources and context.
9.4 Metadata filtering
Under source filtering, metadata filters make sure each touchpoint uses only relevant knowledge, reducing duplication across sources and use cases.
| Condition | Meaning | Allowed |
|---|---|---|
| ALL | Content must contain all selected tags | Once |
| ANY | Content must contain at least one selected tag | Up to five times |
| NONE | Content must contain none of the selected tags | Once |
9.5 The SharePoint Managed Metadata constraint
Genesys retrieves tags from sources such as the File Connector and SharePoint, but for SharePoint only Managed Metadata, from the SharePoint term store, is supported; no other column types are. If you categorise documents with columns such as Choice or Location, Genesys recommends migrating them into Managed Metadata.
That is a project, not a setting. Many SharePoint libraries use Choice columns, and building a term store is a SharePoint governance exercise with its own owner and lead time. Raise it in discovery, because it can move a fabric deployment by months.
9.6 Filtering by conversation context
Static filters apply to every conversation; dynamic filtering adds a filter that changes with each one, based on conversation context. The same document can then serve only the relevant sections to each customer, using what the touchpoint already knows, such as product, segment or authenticated identity. The cost is that correctness now depends on that context: a filter built on stale CRM data produces a confident, well-cited answer from the wrong section. The data feeding dynamic filters needs the same scrutiny as the documents.
Genesys has also added enhanced conversational understanding for Agent Copilot knowledge searches, which helps Copilot recognise when a search is needed and use conversation context to improve the query, and lets agents see the query used to retrieve suggestions. Seeing the query turns a wrong suggestion from a mystery into something diagnosable.
10
Where knowledge surfaces
The two platforms reach different touchpoints, and the assistance products that surface knowledge to agents are packaged and constrained differently. This part maps where knowledge appears, and one telephony constraint that removes voice assistance altogether.
10.1 The touchpoint map
| Touchpoint | Workbench V2 | Knowledge fabric |
|---|---|---|
| Agent Copilot | Yes | Yes |
| Virtual Agent | Yes | Yes |
| AI Guides | Yes | Yes |
| Agentic virtual agent | Yes | Yes |
| Bot and digital bot flows | Yes | Only through Virtual Agent |
| Knowledge portal | Yes | No |
| Knowledge app for Messenger | Yes | No |
- The fabric covers the generative touchpoints, not the classic self-service ones. A portal or Messenger FAQ still needs a knowledge base.
- Which touchpoints accept knowledge is a capability question; which are included in a licence is a commercial one (part 13). Keep them separate in proposals.
10.2 Agent Assist and Agent Copilot
Agent Assist is the longer-standing panel in the agent workspace. Agents type a question or keyword in its search box, and it uses type-ahead to suggest answers as they type. Genesys's availability page says knowledge for manual search and canned responses is included in all tiers, while automatic knowledge retrieval during a conversation needs Agent Assist, and real-time transcription and summarisation need Agent Assist for voice or omnichannel. Manual search suits title and phrase retrieval, which is why Genesys advises against content search for those users.
Agent Copilot is the AI assistant that follows the conversation, surfaces knowledge and can generate answers, and it is where the confidence threshold and enhanced conversational understanding live. For assurance, Genesys Cloud Copilot configuration has an Audit Log tab listing every change to its configuration: the entity, the action (add, enable, edit or disable), who and when.
10.3 The BYOC Premises gap
One constraint can invalidate an architecture, and it sits in the Agent Assist availability page: with BYOC Premises, Agent Assist for voice is not available, because transcription happens after the recording completes and uploads, so real-time summarisation, knowledge surfacing and transcription are not possible. There is no live audio stream to work from.
Our telephony handbook explains why an organisation chooses BYOC Premises: survivability, local media or regulation. This handbook explains why the same organisation then cannot surface knowledge to voice agents in real time. The two requirements conflict, and somebody must be told before build. What still works is manual knowledge search and the post-call path, with transcription and analysis after upload. Genesys also notes that Agent Assist currently supports only English for outbound conversations.
11
Bot flows, the knowledge portal and canned responses
Knowledge reaches customers through bot flows and the knowledge portal, and agents also use canned responses, a controlled sibling of knowledge that is often governed by a different team.
11.1 Knowledge in bot flows
Workbench knowledge bases are linked to a flow with or without Virtual Agent; fabric configurations are linked to flows with Virtual Agent enabled. Knowledge settings sit in the Knowledge view of Architect's Natural Language Understanding menu. Response bias weights interpretation towards intents, neutral or knowledge. The content type can be Standard (the full matching article), Highlight (the answer highlighted within the article) or Generative (an AI-generated answer); Standard is not available with a fabric configuration. Disambiguation can handle up to 20 possible answers, returning three by default.
11.2 Taking control with a reusable knowledge task
Since April 2024, flow authors can designate a reusable task that runs immediately after the bot finds one or more matching articles, deciding when to show the article, disambiguate or ask for feedback; the actions that trigger it are Wait for Input and Digital Menu. That addresses the biggest quality problem in FAQ bots, which is speaking the best match whatever its confidence. Separately, the Find KnowledgeBaseDocument by ID action looks up an article directly, which Genesys says reduces memory use and removes the need for switch statements.
11.3 The knowledge portal
The knowledge portal presents articles to customers, who can search, browse by category and start a conversation with a bot or an agent. It is set up under Messenger Configurations: choose a knowledge base and categories, then copy the deployment snippet to your website. Options include the welcome or fallback message, colours, a header image, top viewed articles, categories with icons, and the ability to disable the portal.
Two consequences follow. The portal is a Messenger configuration, not a separate product, so its availability travels with Messenger. And it is embedded on a site the organisation controls, so content security policy, styling and cookie consent sit with the web team, who are rarely invited to contact centre design reviews. Because customers browse categories, the one-category rule in part 5 is customer-facing.
11.4 Canned responses: knowledge's controlled sibling
Canned responses are libraries of standard text that agents read out or insert into chats, emails and messages. They support substitutions such as the customer's and agent's names, attachments, and response types including email templates and signatures, WhatsApp and SMS templates and forms. Forms let an agent or bot send a structured request, such as a list picker, date picker or free text, in a messaging interaction and receive the reply.
Agents with the Genesys Cloud User role can view and use libraries; authors need the Responses permissions. Since 31 August 2026, libraries can be restricted by division. Genesys's troubleshooting list for an agent who cannot see a library is a good general diagnosis: is the library in a division the user can access, does the user have the permissions, and does queue segmentation include the library for the current interaction?
12
The knowledge optimizer, and what it cannot tell you
The optimizer is what lets a knowledge base be operated rather than launched and left. Its limits matter as much as its features when you design assurance.
12.1 What the dashboard shows
For knowledge bases, the optimizer dashboard reports all queries in a period split into answered and unanswered, answered and unanswered queries by originating application, the top 50 articles with frequency and feedback, and the top 20 answered and unanswered queries. You can filter by touchpoint: knowledge portal, knowledge app for Messenger, Dialog Engine and digital bot flows and Agent Assist, or Agent Copilot for a fabric configuration; selecting several sources of one type shows aggregate figures. From an article tile, Article View shows its details, feedback, views and the other queries it answered.
12.2 The optimise loop
The optimise view is the working tool. Unanswered queries appear as tiles ranked by high, medium and low impact, each showing how often customers asked it or something similar. Opening one shows similar phrases and suggested related articles that might hold the answer. From there you attach the phrase to an existing article or create a new article with the phrase as its question, saving it as a draft for an author.
Two mechanics matter: unanswered queries are clustered overnight, so today's testing appears tomorrow; and the view covers up to the last 10 days.
12.3 Its limits
- Ten days. The optimise view covers up to the last 10 days, so monthly or quarterly reviews cannot rely on it alone.
- Fabric reporting is thinner: data for a knowledge fabric configuration is available for the last 10 days only, covering interactions with generated answers, feedback submissions and answer-generation queries with negative feedback.
- Feedback is optional, so it shows the customers who rated an answer, not those who gave up.
- It is a period snapshot, not a trend across months.
12.4 Keeping your own record
For a fabric deployment the consequence is uncomfortable: a generated answer can be wrong, and the only signal may be a customer who chose to give negative feedback. Ten days of feedback counts is not an audit trail. If an organisation will let AI answer customers from documents, it needs its own record of queries, retrieved sources and answers, captured at the time. Genesys's data retention page lists export routes, including the Amazon EventBridge integration for customer audit data, alongside the notification and analytics APIs.
13
Licensing, languages and permissions
Three practical questions decide what a design can promise: what the licence includes, which languages are supported where, and how access to knowledge is controlled. QVCCS sells no licences and quotes no prices; our Experience Units guide explains how AI consumption is metered.
13.1 What the plans include
On Genesys's summarised plan comparison (dated 3 November 2026 when we checked), the knowledge workbench and knowledge optimizer are included in Genesys Cloud CX 2, CX 3 and CX 4, and available on CX 1 through Experience Units. Agent Copilot capabilities, including automatic knowledge surfacing and answer highlighting, are included in CX 4 and use Experience Units on CX 1 to CX 3. Fabric setup also needs specific licences and permissions, and answer generation is AI consumption, so treat it as an Experience Units question even where the platform is included.
13.2 Languages
- Genesys's bots documentation says the knowledge workbench supports language search in English and German.
- Content search follows the workbench V2 entry on Genesys's supported languages page, and the threshold guidance singles out Filipino, Greek, Hebrew and Malay for testing above 0.7.
- The fabric is released language by language; in July 2026, for example, Genesys added Czech and Hungarian for Agent Copilot and all Virtual Agent tiers.
- Generated answers use the touchpoint's configured language, and an English source with a non-English Virtual Agent is not officially supported.
- Agent Assist currently supports only English for outbound conversations.
13.3 The permission domain
Counted from Genesys's published API definition on 15 November 2026, the knowledge domain has 75 permissions across 130 operations under the knowledge API, plus analytics:knowledgeAggregate:view for reporting. They group cleanly, and the action sets are not uniform: export jobs have add, delete and view but no edit; synchronisation has add, edit, view, upload and deleteFile but no plain delete; and feedback uses create rather than add.
| Group | Permissions | Controls |
|---|---|---|
| Knowledge bases | knowledge:knowledgebase add, edit, delete, view, search | Whole bases |
| Articles and versions | knowledge:document add, edit, delete, view, upload; documentVersion add, view | Article content and history |
| Taxonomy | knowledge:category and knowledge:label add, edit, delete, view | Categories and labels |
| Import and export | knowledge:importJob add, edit, delete, view; exportJob add, delete, view | Bulk content changes |
| Connectors | knowledge:connection and knowledge:integrationSource add, edit, delete, view | External article sources |
| Fabric sources | knowledge:source add, edit, delete, view; syncJob and synchronization permissions | Ingestion and synchronisation |
| Settings and preview | knowledge:knowledgeSetting add, edit, delete, view, search, searchPreview | Configurations, including preview |
| Search and feedback | knowledge:search:edit; knowledge:feedback create, edit, view | Matching behaviour and ratings |
13.4 Notes for role design
- knowledge:search:edit changes how every article retrieves, so it belongs with the knowledge manager, not with authors.
- knowledge:knowledgeSetting:searchPreview gives access to preview; give it to authors generously, because a team that can test does not need to experiment on live customers.
- Import, export and source permissions are bulk-change capabilities; keep them with one or two named platform roles.
- analytics:knowledgeAggregate:view is not on Genesys's hourly interacting disqualifying list (we checked the list on 15 November 2026), but other aggregate analytics permissions are, so check any reporting role against the list before giving it to hourly users.
14
Governance, programme design and what goes wrong
Most failures in knowledge programmes come back to governance. Genesys provides a lifecycle, revision history, feedback, an optimizer and a loop from unanswered query to new article. It does not provide an owner, and knowledge degrades silently: nothing alerts when an answer becomes wrong, and customers who get one mostly just call back.
14.1 Three named roles
- A content owner for each domain: a subject-matter expert, in Genesys's words well versed in the subject, not a project consultant who leaves in twelve weeks.
- A knowledge manager who owns search quality: thresholds, phrase distribution, categories and the weekly optimise loop.
- A publisher who controls what reaches customers: publishing, bulk operations, import and export, deliberately separate from authors.
14.2 The fabric sharpens governance
With the fabric, the corpus is documents rather than reviewed articles, and anyone who can write to an ingested folder is publishing to customers. The mitigations are ordinary but need agreeing with SharePoint and information governance teams who are usually outside the programme: a dedicated knowledge library, restricted write access, required metadata and a review state for documents.
One question must be answered explicitly in every fabric design: who is accountable when a generated answer is wrong? With a curated article, it is the article's owner. With an answer built from part of a document whose owner may be unknown, accountability disappears unless someone is named. Name the role, give them the export path from part 12, and agree the response to a challenged answer: narrow the filter or remove the source, fix the content, retest in the preview widget, then restore.
14.3 Designing the programme
| Question | If yes | If no |
|---|---|---|
| Is the main consumer a bot flow without Virtual Agent? | Workbench V2 for that channel | The fabric is available for Agent Copilot, Virtual Agent, AI Guides and agentic virtual agents |
| Is the source structured CRM articles? | Article connectors, with the source kept authoritative | Consider the fabric |
| Is the content mainly unstructured documents? | The fabric, after a chunk-aware content review | Curated question-and-answer articles |
| Does the organisation use SharePoint Managed Metadata? | Fabric filtering works as designed | Plan a term store project first |
| Must exact approved wording be used? | Canned responses, scoped by division | Knowledge |
- Then fix the constraints before writing content: content search per base, the category hierarchy, one language per base, thresholds from the published defaults, and the telephony check for voice assistance.
- Build one domain and one touchpoint end to end, through to the optimise loop, before widening.
- Measure before scaling: twenty real phrasings through the actual flow, the answered and unanswered ratio, and negative feedback on generated answers.
14.4 Fourteen things that go wrong
| Symptom | Likely cause | Fix |
|---|---|---|
| Confident wrong answers | The highest-confidence article is not the right one, or the threshold is too low | Test real phrasings and use Genesys's defaults by base type (part 3) |
| Wrong answers after a migration | The base moved to content search but kept the 0.5 threshold | Retune to 0.65, or 0.7 if not English; 0.65 or below with generated answers (part 3) |
| Fragments with no context | Fixed 400-token chunking split the answer | Self-contained sections, headings and key facts first (part 8) |
| A table or diagram is ignored | Only text is indexed | Describe it in the surrounding text (part 8) |
| Phrases seem to do nothing | The base uses content search | Follow one model per base (parts 2 and 4) |
| Content search cannot be enabled | It is set only at creation | Export, create a new base and import, then relink touchpoints (part 4) |
| ServiceNow articles arrive malformed | Custom templates are not supported | Confirm standard templates in discovery (part 7) |
| A fabric source is missing from the configuration list | The source is not yet indexed | Check ingestion status first (part 9) |
| A filter matches nothing or everything | SharePoint Choice or Location columns; only Managed Metadata is read | Move categorisation into the term store (part 9) |
| Answers in the wrong language | A non-English Virtual Agent on an English source | One language per base or configuration (part 13) |
| Agents see the wrong variation | The rightmost variation tab wins | Order tabs deliberately and test (part 5) |
| Live content disappeared after a bulk change | Bulk restore returned unpublished articles to unpublished | Count published articles first and rehearse on a subset (part 6) |
| Voice agents get no suggestions | BYOC Premises: Agent Assist for voice is unavailable | Manual search and post-call analysis; decide with telephony (part 10) |
| Knowledge went stale within a quarter | No owner, review cadence or optimise loop | Name the three roles and run the weekly loop (part 14) |
Questions
Common questions
What is the difference between the Genesys knowledge workbench and knowledge fabric?
The workbench holds structured question-and-answer articles in Genesys Cloud, authored or copied from Salesforce, ServiceNow or the universal connector, and serves bot flows, the knowledge portal, Messenger and AI touchpoints. The fabric connects document sources such as SharePoint, divides text into chunks and generates cited answers for Agent Copilot, Virtual Agent, AI Guides and agentic virtual agents, without copying documents in.
Can the knowledge fabric be used in Genesys Cloud bot flows?
Genesys's comparison says the knowledge fabric is not currently supported in bot and digital bot flows, while its Architect guidance links a fabric configuration to flows with Virtual Agent enabled. In practice, classic bot flows use workbench knowledge bases, and the fabric reaches flows through Virtual Agent. The knowledge portal and Messenger knowledge app also need workbench bases.
What confidence threshold should I use for Genesys knowledge?
Genesys recommends 0.5 for knowledge bases without content search, 0.65 for English content search bases and 0.7 for other languages, with testing above 0.7 suggested for Filipino, Greek, Hebrew and Malay. If you use AI-generated answers, Genesys advises 0.65 or below for all languages. Retest whenever a base's content model or language changes.
Can I turn on content search for an existing knowledge base?
No. Genesys states that content search cannot be enabled automatically for existing articles. You export the articles, create a new knowledge base with content search enabled and import them, then relink every touchpoint. Content search caps phrases at 100 per article, and Genesys advises removing phrasings from content search bases altogether.
How does the knowledge fabric split documents?
Genesys states that all ingested content is divided into fixed chunks of 400 tokens, whatever the document's structure, and that only text is indexed. The top three chunks are returned by default and used with the query to generate a cited answer. Writing self-contained sections with headings, putting key points first and describing tables in words improves results.
Does Agent Assist work with BYOC Premises?
Not for voice. Genesys states that with BYOC Premises, Agent Assist for voice is not available, because transcription happens only after the recording completes and uploads, so real-time summarisation, knowledge surfacing and transcription are not possible. Manual knowledge search and post-call analysis still work.
Sources
- Control knowledge behavior in your voice or digital bot flow – Genesys Cloud Resource Centerhelp.genesys.cloud
- Genesys Cloud supported languages – Genesys Cloud Resource Centerhelp.genesys.cloud
- Import or export articles in a knowledge base v2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Deprecation: Knowledge workbench V1 – Genesys Cloud Resource Centerhelp.genesys.cloud
- About bots – Genesys Cloud Resource Centerhelp.genesys.cloud
- About canned responses – Genesys Cloud Resource Centerhelp.genesys.cloud
- About the knowledge portal – Genesys Cloud Resource Centerhelp.genesys.cloud
- About the knowledge fabric – Genesys Cloud Resource Centerhelp.genesys.cloud
- About the knowledge workbench V2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add a category to a knowledge base V2 article – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add a third-party knowledge base to the knowledge workbench – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add a touchpoint or segment for a knowledge base article – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add categories to a knowledge base V2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add knowledge to your bot flow – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add phrases to a knowledge base V2 article – Genesys Cloud Resource Centerhelp.genesys.cloud
- Add question and answer articles to a knowledge base V2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Browse by categories in the knowledge portal – Genesys Cloud Resource Centerhelp.genesys.cloud
- Built-in connectors for knowledge articles – Genesys Cloud Resource Centerhelp.genesys.cloud
- Canned responses forms quick start guide – Genesys Cloud Resource Centerhelp.genesys.cloud
- Canned responses overview – Genesys Cloud Resource Centerhelp.genesys.cloud
- Configure Genesys Cloud Copilot – Genesys Cloud Resource Centerhelp.genesys.cloud
- Configure Genesys Agent Copilot settings – Genesys Cloud Resource Centerhelp.genesys.cloud
- Connect guides to a knowledge base or configuration – Genesys Cloud Resource Centerhelp.genesys.cloud
- Create a knowledge configuration – Genesys Cloud Resource Centerhelp.genesys.cloud
- Data retention periods and export functionality by feature – Genesys Cloud Resource Centerhelp.genesys.cloud
- Divisions for Canned Responses - best practices and troubleshooting – Genesys Cloud Resource Centerhelp.genesys.cloud
- Edit multiple knowledge base articles at once – Genesys Cloud Resource Centerhelp.genesys.cloud
- Filter knowledge dynamically by conversation context – Genesys Cloud Resource Centerhelp.genesys.cloud
- Genesys Agent Assist availability in Genesys Cloud licenses – Genesys Cloud Resource Centerhelp.genesys.cloud
- Surface knowledge manually with Genesys Agent Assist – Genesys Cloud Resource Centerhelp.genesys.cloud
- Include or exclude articles from knowledge surfacing – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge base content search overview – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge fabric: key concepts and content strategies – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge fabric overview – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge optimizer dashboard – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge workbench overview V2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge workbench quick start guide V2 – Genesys Cloud Resource Centerhelp.genesys.cloud
- Knowledge workbench V2 supported configuration – Genesys Cloud Resource Centerhelp.genesys.cloud
- Manage knowledge configurations – Genesys Cloud Resource Centerhelp.genesys.cloud
- Optimize your knowledge base content – Genesys Cloud Resource Centerhelp.genesys.cloud
- Understand confidence levels – Genesys Cloud Resource Centerhelp.genesys.cloud
- Understand knowledge base V2 categories and labels – Genesys Cloud Resource Centerhelp.genesys.cloud
- Understand knowledge base V2 questions and answer articles – Genesys Cloud Resource Centerhelp.genesys.cloud
- View and restore previous versions of a knowledge base article – Genesys Cloud Resource Centerhelp.genesys.cloud
- What types of knowledge connectors are available and when should they be used? – Genesys Cloud Resource Centerhelp.genesys.cloud
- April 29, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- August 18, 2021 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- August 24, 2022 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- August 31, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- February 14, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- February 22, 2023 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- January 31, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- July 13, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- November 16, 2022 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- October 28, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- September 9, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
- Genesys Cloud Platform API v2 definition (permission requirements per operation) – Genesysapi.mypurecloud.com
- Hourly interacting license disqualifying permissions – Genesys Cloud Resource Centerhelp.genesys.cloud
- Genesys Cloud CX summarised plan comparison – Genesyslibrary.genesys.com
Keep reading
More guides
AI agent, bot or copilot? Choosing the right AI for each contact centre journey
Scripted bots, virtual agents, agentic AI agents and copilots on Genesys Cloud: who each serves, what it does well, how it is charged and how to combine them.
Read the guideVoice AI cost calculator for Genesys Cloud: native and third-party compared
Compare the list-price cost of Genesys bot flows, virtual agents and agentic agents with Google, Amazon, Deepgram and ElevenLabs voice AI, per resolved call.
Open the calculatorNative or third-party voice AI on Genesys Cloud: how to choose
Genesys bot flows, virtual agents and agentic agents, or Dialogflow, Lex, Deepgram, ElevenLabs and others? How each connects, costs and is governed.
Read the guide