Back to all articlesAI knowledge preparation

Separate Current Rules From Historical Rules in AI Knowledge

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A service policy published in June may not take effect until July. A document published in August may clarify the policy that applied in May. If an AI assistant treats publication date as the deciding factor, it can give a well-sourced answer that is wrong for the customer's situation.

To handle effective dates in AI knowledge, record each rule's publication date separately from its effective period, define which business event determines applicability, and retrieve against that event's date. Preserve earlier versions for historical questions. When applicability is unknown or conflicting, make that uncertainty visible instead of assuming that the newest document wins.

The practical starting point is a temporal knowledge worksheet: one record per rule version and scope, with enough information to explain why that version applies. The filled fictional worksheet below shows how to build it and test it before relying on it for operational answers.

1. Decide which date the question actually needs

“What is our policy?” and “What policy applied to this booking?” are different retrieval tasks. The first usually asks for the rule effective at the time of the question. The second may depend on when the booking was confirmed, when service occurred, or another explicitly defined event.

Before preparing documents, identify the date-bearing event for each rule family. Do not let the assistant choose one from conversational convenience.

  • Current operational question: What rule should staff use for a new transaction now?
  • Historical applicability question: What rule applies to a transaction associated with an earlier event?
  • Scheduled-change question: What rule will apply after a confirmed future effective date?
  • Historical knowledge question: What information had been published or recorded by a particular date?

The last two historical concepts are easy to confuse. A correction published later might establish what rule actually applied earlier, but it cannot establish what staff could have read before that correction existed.

A useful intake question is: Are we answering what applied then, or what was known then? Keep those modes separate. If the user asks about an existing transaction without providing the relevant event date, request it or retrieve it through an authorised workflow. Today's date is not a substitute.

2. Give every version an explicit temporal record

A publication date is evidence about a document's release, not automatically evidence about its operational start. Use separate fields so retrieval does not have to infer that distinction from prose.

  • Rule family and version ID: Stable identifiers connecting successive versions without overwriting them.
  • Source locator: The document and section supporting the record.
  • Publication date: When the source was published, if established.
  • Effective date: When this version starts applying to its defined scope.
  • Expiry date: When this version stops applying, if established.
  • Applicable period: A plain-language explanation of the interval and the event it governs.
  • Scope and exceptions: Service type, customer group, channel or other conditions that change applicability.
  • Verification status: Whether an accountable owner has confirmed the interpretation.

Also keep a recorded-at timestamp: when the knowledge system received this information. That helps distinguish delayed ingestion from delayed publication and supports investigations into past answers.

Choose an interval convention and document it. A practical option is an inclusive start and exclusive end: the version applies from its effective date up to, but not including, its expiry date. Adjacent versions can then meet at the same boundary without both applying.

Use an explicit timezone when time-of-day boundaries matter. For date-only sources, establish the business interpretation with the source owner rather than quietly inventing midnight or a timezone.

3. Filled fictional worksheet: successive service policies

Fictional worked example: The following organisation, policies, dates and service terms are hypothetical. They illustrate a recommended preparation design, not an existing product feature or customer result.

A fictional equipment-maintenance business changes how far ahead customers must request a service-visit reschedule. Its source owner confirms that the relevant event is booking confirmation, not the visit date or the date of the reschedule request. Existing confirmed bookings retain their original terms.

The hypothetical worksheet uses date-only boundaries in the business's agreed operating timezone. Expiry dates are exclusive. “Open” means no confirmed end date is recorded; it does not mean the rule can never change.

RULE FAMILY: Standard service-visit rescheduling
SCOPE: Standard maintenance bookings only
APPLICABILITY EVENT: Booking confirmation date
EXCLUDED SCOPE: Priority support contracts
SOURCE OWNER: Service operations owner
INTERVAL CONVENTION: Effective date inclusive; expiry date exclusive

VERSION: SR-01
SOURCE: Fictional Service Handbook A, rescheduling section
PUBLICATION DATE: 2025-01-10
EFFECTIVE DATE: 2025-02-01
EXPIRY DATE: 2025-05-01
APPLICABLE PERIOD: Bookings confirmed from 2025-02-01
                  through 2025-04-30
RULE: Request a reschedule at least 2 business days
      before the scheduled visit.
HISTORICAL APPLICABILITY: Still governs those bookings
                          after this version expires.
SUPERSEDED BY: SR-02 for new confirmations only
VERIFICATION: Confirmed by fictional source owner

VERSION: SR-02
SOURCE: Fictional Service Update B, rescheduling section
PUBLICATION DATE: 2025-04-15
EFFECTIVE DATE: 2025-05-01
EXPIRY DATE: 2025-09-01
APPLICABLE PERIOD: Bookings confirmed from 2025-05-01
                  through 2025-08-31
RULE: Request a reschedule at least 3 business days
      before the scheduled visit.
HISTORICAL APPLICABILITY: Still governs those bookings
                          after this version expires.
SUPERSEDED BY: SR-03 for new confirmations only
VERIFICATION: Confirmed by fictional source owner

VERSION: SR-03
SOURCE: Fictional Service Update C, rescheduling section
PUBLICATION DATE: 2025-08-12
EFFECTIVE DATE: 2025-09-01
EXPIRY DATE: Open; no confirmed end date recorded
APPLICABLE PERIOD: Bookings confirmed on or after 2025-09-01
RULE: Request a reschedule at least 1 business day
      before the scheduled visit.
HISTORICAL APPLICABILITY: Does not replace the terms
                          of earlier confirmed bookings.
SUPERSEDED BY: None recorded
VERIFICATION: Confirmed by fictional source owner

These records make a critical distinction: expiry ends eligibility for new booking confirmations. It does not erase the rule's applicability to an existing booking. A simple “expired” filter would lose that distinction.

The worksheet also exposes a coverage limit. It establishes nothing about bookings confirmed before the first effective date. Retrieval should flag that gap rather than extend the earliest available rule backwards.

4. Prepare chunks without stripping away applicability

Even a good worksheet can fail if document preparation separates a rule from its temporal context. A chunk containing only “request at least three business days ahead” is unsafe to use without its scope and applicable period.

Keep each retrievable rule unit attached to its version ID, applicability event, period, scope and source reference. If a long document becomes several chunks, carry those fields into every chunk that could independently support an answer. Preserve any exception that materially changes the rule beside it or through an explicit dependency.

Do not assign one effective date to an entire handbook when sections change independently. Split the records by rule family or section. Otherwise, a recently updated contact-details page can make an unchanged service rule appear newly effective.

Maintain both structured metadata and readable applicability text. Structured dates support deterministic filtering; plain-language conditions help reviewers catch mistakes such as confusing booking date with appointment date. Neither replaces the source evidence.

When the source lacks an effective date, mark it unknown. A file's modified date, upload date or retrieval date should not silently fill the gap. Route the record for confirmation and limit its use until the applicability question is resolved.

5. Retrieve for the applicable event, not the newest file

A recommended retrieval design separates applicability checks from relevance ranking. First determine the question's temporal mode and scope. Then identify eligible rule versions. Only then rank supporting passages within that eligible set.

  1. Identify intent. Is the user asking about a new booking now, an existing booking, a future change or what was known earlier?
  2. Resolve the governing event. For the fictional worksheet, use booking confirmation date. Ask for missing information when it changes the answer.
  3. Match scope. Exclude priority support contracts from this rule family rather than treating them as ordinary bookings.
  4. Apply the interval. Select the version whose start includes the event date and whose exclusive end follows it.
  5. Check exceptions and conflicts. Do not let semantic similarity settle competing applicability claims.
  6. Return the rule with its basis. State the relevant date and version in language the reader can inspect.

For example, in the hypothetical scenario, a booking confirmed on 2025-04-28 and scheduled for 2025-09-10 remains under SR-01. The later visit date does not move it to SR-03. Conversely, a new booking confirmed on 2025-09-10 uses SR-03.

For a future question asked before SR-03 takes effect, label that version as scheduled rather than current. For a “what was known then?” question, restrict evidence by publication or recorded-at time as appropriate, and explain which history is available. A date-only publication record may not support a precise answer about an hour within that day.

6. Test boundaries, missing dates and misleading freshness

Use a fixed evaluation date instead of letting tests depend on the day they run. The following tests are hypothetical and use the worksheet above, with an evaluation date of 2025-09-10 unless stated otherwise.

  • Current new booking: “What notice applies to a standard booking confirmed today?” Expected: SR-03 and one business day.
  • Historical booking: “The booking was confirmed on 2025-04-28. What notice applies?” Expected: SR-01, even if the visit is later.
  • Exact boundary: “Confirmed on 2025-05-01.” Expected: SR-02, not SR-01.
  • Second boundary: “Confirmed on 2025-09-01.” Expected: SR-03, not both SR-02 and SR-03.
  • Future publication: At an evaluation date of 2025-08-20, ask about a new booking that day. Expected: SR-02; SR-03 is published but not yet effective.
  • Missing event date: “Can I move my existing visit?” Expected: request the booking confirmation date before choosing a version.
  • Wrong scope: “What applies to my priority support booking?” Expected: identify the scope mismatch, not reuse the standard rule.
  • Coverage gap: “Confirmed on 2025-01-20.” Expected: insufficient historical coverage, not SR-01 by default.

Add a deliberately newer document that merely summarises an old version without replacing it. Confirm that publication freshness does not override the explicit applicable period. This tests the failure the worksheet is designed to prevent.

Score version selection separately from answer wording. A polished answer using the wrong version is still a retrieval failure. Record the selected version, event date used, scope decision, source reference and whether clarification was requested.

7. Recover from conflicts without rewriting history

Overlapping periods can represent a data error, a real scope difference or an unresolved source conflict. Do not automatically shorten one period just to make the records fit. Check the underlying conditions and ask the accountable owner to resolve genuinely competing rules.

Likewise, do not fill a gap by carrying the previous version forward unless the source supports continuity. Mark the gap and make the assistant explain that the applicable rule cannot yet be established.

When correcting metadata, retain the prior record and the reason for the correction. Distinguish a source revision from a preparation error: they have different implications for reviewing earlier answers. If the wrong period was indexed, rebuild the affected retrieval records and rerun relevant historical and boundary tests.

There is a trade-off between keeping every version available and making current retrieval simple. Use a current-view collection or filter if helpful, but retain a history path for earlier transactions. Do not make archival placement synonymous with “never applicable.” Where filtering capabilities are limited, explicit temporal text can help, but it is a weaker safeguard than a tested eligibility check.

8. Start with one rule family and inspect the evidence

Choose one service policy with successive versions and a clear owner. Complete the worksheet, confirm the governing event, and run the hypothetical test patterns against your own verified dates and scopes. This is a preparation and evaluation task; it does not require assuming that any particular platform already provides these controls.

Track three observable outcomes: whether the expected version was selected, whether unresolved cases triggered clarification, and whether the answer exposed its applicability basis. Review failures by cause: missing metadata, incorrect interval logic, lost scope, or answer generation ignoring the retrieved record.

A useful completion criterion is that every test has an expected outcome, every wrong-version selection has been corrected or blocked from operational use, and every unresolved interval has a named owner. That gives the next editor or operator something more reliable than a folder sorted by publication date: a traceable explanation of which rule applies, to which event, and why.

AI knowledgeEffective datesRetrieval testingService policies

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications