How to update and test an agent's knowledge base

Knowledge-base inventory: source, validity and owner

Inventory the information guiding service: its source document, supported rule, agent using it, owner and effective date. Hours, channels and service conditions can change at different times. Identify the changed fact before uploading files so all affected configurations and questions can be located.

Large unowned collections make current material difficult to identify. Prefer focused documents and an approved origin for each rule.

Review content and selection

Upload the updated version to Tigy's knowledge base and wait for processing. Check attached documents and deselect versions contradicting the current rule.

Uploading a workspace file does not automatically select it for every agent. Check each agent needing that information.

Repeat questions revealing the difference

Test the new rule, an exception and a question mentioning the previous condition. Compare answers with the current document and check clarification for missing information.

Include common questions phrased differently. Repeating the file title does not represent every service query.

Close the change with evidence

Record updated material, checked agents and repeated scenarios. Check the published version for configuration changes and observe new conversations.

Documents do not replace live individual lookups. Maintain policies editorially and use integrations for changing status, balances or availability.

Test before-and-after policy conditions

For fictional rescheduling changes, test new deadlines, boundary cases and specialist exceptions.

Include old terminology without accepting outdated assumptions.

Verify affected languages and branches independently.

Retire content as carefully as you introduce it

When a service is discontinued, document revision requires more than deleting its heading. Search frequently asked questions, attachments, instructions, and service examples for references. Content can continue directing callers toward a retired channel after the main description has disappeared. Retirement tests should ask using both the old service name and an informal description, verifying that the agent presents only the approved alternative.

If customers still hold old documents, provide a brief explanation of the change and a real next step. Do not frame the question as the caller's mistake. Knowledge may retain enough information to recognize the old name without presenting the previous rule as a valid option. Preserve editorial history outside conversational sources and connect retirement to relevant regression cases. This lets the agent explain the change without accidentally reviving superseded guidance. Also inspect attached sources after replacing a file: a correct new document does not help if an older duplicate remains selected and answers the same question differently. Completing retirement therefore includes checking the actual source collection, not just the file the editor most recently changed.

Inventory facts that can change

Knowledge review starts by identifying which decisions depend on each document. A service file may contain hours, addresses, contact channels, and collection conditions. Those facts have different owners and change at different rates. If the entire file is merely labeled updated manual, reviewers may improve wording while leaving an old address in a rarely read section. The inventory should connect each important fact to an accountable person and an approved source.

Separate institutional facts, policies, and individual information. A location's address and a cancellation policy can be shared material when approved. A particular order's status requires authorized, current system access. Uploading an order export to avoid integration creates rapidly aging content and mixes personal information with reusable knowledge. Resolve that confusion before attempting to improve answer quality.

For each document, record its purpose and the questions it should answer. A service guide may explain required documents and contact routes; it should not guarantee individual appointment availability. This brief helps identify inference beyond the source. A statement that review usually takes several days does not confirm the deadline for one specific case. Content should distinguish general information from individual outcomes so conversation can maintain that same distinction.

Look for duplication too. A policy repeated in instructions, a PDF, and an internal page may be updated on different dates. Choose the source the team can maintain and remove or coordinate other copies. Knowledge quality does not improve merely by accumulating documents. It improves when sources provide consistent answers and have a known review route. In Tigy, check both file processing and association with the agent: correct material in the workspace does not prove the tested conversation consults it. Record the reviewed file identity so later replacement cannot be mistaken for the same approved source.

Resolve conflicts before testing retrieval

An answer using the wrong passage may indicate retrieval trouble, but it may also reveal conflicting documents. Before adjusting instructions, locate the fact across all selected sources. An old policy may allow cancellation by telephone while a newer document requires a different channel. If both remain available without context, there is no single dependable answer the agent can select merely through better wording.

When a conflict appears, identify effective dates, audience, and service. Different rules may legitimately apply to different products. Add enough context to distinguish those cases and require clarification when necessary. If one rule truly replaced another, remove the superseded source from the agent's selected knowledge and preserve history elsewhere for the team. Do not expect conversation to infer precedence because one file has a newer-sounding name.

Review implicit references. Phrases such as as agreed, under the above conditions, and during this period may lose meaning when a passage is read independently. Rewrite relevant information with an explicit subject, condition, and consequence. A paragraph should identify its service and exceptions. This also helps human review: someone uninvolved in drafting can verify the rule without reconstructing the entire original document.

Preserve conditions when information is conditional. Do not reduce a modality-specific price table to one amount or remove an important exception merely to shorten answers. Conversation may summarize and ask which situation applies, but the source must support that choice. After correction, test both a direct question and one suggesting an outdated premise. Ask for the channel that is no longer accepted, for example. Expected behavior is to correct the premise using the current source and describe the real alternative instead of agreeing with the caller's suggestion. Include an ambiguous product reference as well, confirming that the agent asks the necessary distinction before selecting a rule.

Assess coverage, absence, and conversational use

A knowledge test needs both answerable and unanswerable questions. Select facts distributed across the document, exceptional conditions, and an absent fact. If every test copies headings literally, the result measures unusually easy input. Use audience language: familiar abbreviations, incomplete references, and descriptions of situations. A caller may ask whether everything can be completed today while the document describes separate requirements and stages.

For each question, write expected content rather than an exact sentence requiring word-for-word reproduction. Identify mandatory facts, conditions that cannot disappear, and the next step that must remain clear. Different wording may still be correct. Conversely, wording resembling the source may be wrong if it omits the condition changing the decision. This distinction helps reviewers assess meaning instead of superficial similarity.

Test source combinations too. A location-specific inquiry may require hours from one document and an address from another. Verify that both facts belong to the same location and their combination does not invent a policy. When the caller corrects a detail, the agent should update the answer. This matters particularly for similar service or place names, where fluent responses can conceal incorrect associations.

Move important cases to voice. A deadline with several qualifications may be accurate in text but difficult to follow aloud. The agent can first state the applicable rule and then explain the relevant exception. If it repeats an entire long list unnecessarily, the source and presentation instructions may need review. The goal is not to reduce every document to tiny sentences; it is to ensure approved content is used understandably. Preserve failures through fictional examples for repetition after updates, including the unanswerable case that must continue producing an honest alternative. Track whether a failure concerns missing knowledge, wrong retrieval, or poor presentation so the correction targets the right layer rather than adding redundant content.

Define what happens when a policy changes

Maintenance should begin when a policy owner approves a change. The team needs to know which document will change, which agents use it, and which regression questions must run again. Waiting for the next general review can leave incorrect information circulating. Conversely, updating a file without checking processing and source selection can create a false sense of completion.

Keep a short change record: previous fact, new fact, effective date, approving owner, and affected examples. That record need not enter the agent's knowledge; it helps the team understand the revision. If a policy starts next month, distinguish current rules from future rules clearly. Test one question about today and another about a later date to verify that answers respect effective dates.

When a change affects previously communicated commitments, agree on handling with the responsible operation. Updating knowledge does not independently resolve earlier promises or update external systems. The agent should explain only the approved procedure for such cases. Before closing the revision, verify the selected source, execute affected examples, and retain response evidence. Maintenance then ends with applied, verified content rather than merely a saved file. Also review the questions that remain unanswered: a policy update can introduce a new exception without providing the information needed to explain it. Route that gap to the content owner instead of encouraging the agent to improvise.