AI uses for records
Business retrieval temporal filters: scoring results from the wrong quarter
By SourceX Editorial · Updated
10 sections
Short answer
Treat a relevant passage from the wrong quarter as a retrieval miss. Resolve each question's period into explicit dates using the as-of date and the fiscal calendar, label the controlling source for that period, and include same-template documents from adjacent periods as hard negatives. Score period-correct recall, wrong-period intrusion and superseded versions, then check the answer's stated period, citation support and abstention.
Retrieval
Business retrieval temporal
Key takeaways
- A relevant passage from the wrong quarter is a hard negative for period questions, not a partial match.
- Resolve relative periods into explicit dates using the as-of date and the company's fiscal calendar before filtering.
- Filter on the date field that matches the question: period covered, approval, effective range or restatement.
- Time zones, partial dates and rounding rules shift quarter edges; test boundaries explicitly.
- Report wrong-period intrusion and superseded-version rates alongside precision and recall.
- Include trend, current-state and unclosed-period questions with their own expected behavior.
How do you evaluate retrieval of a relevant result from the wrong quarter?
Score it as a miss. In business retrieval, the period a question asks about is a hard constraint, so a passage that answers the right question for a different quarter is a hard negative, however similar it looks. Build each test item with the question's as-of date, the resolved period and its basis (fiscal or calendar), the controlling source for that period, and the adjacent-period documents a retriever is likely to confuse with it.
Then measure two things separately. Retrieval: did the controlling passage for the right period come back, and how many wrong-period passages came back with it. Answer: did the response state the right period, cite the controlling source, and decline when the period is not closed or the user cannot see the evidence. A system that scores well on topical relevance and badly on these checks will give confident answers about the wrong quarter.
Why can't similarity search tell one quarter from the next?
Business documents repeat themselves by design. A quarterly renewal review, a board pack or a close checklist uses the same template every period, so the second quarter's memo and the third quarter's memo share almost every word except the dates and the numbers. Embedding similarity rewards exactly that shared wording, which means the nearest neighbor of a question about one quarter is very often the same report for another quarter.
Research on time-sensitive question answering makes the same point at the question level. The TimeQA dataset treats a question as time-sensitive when it contains a time specifier, such as in 2007 or before 2010, and changing that specifier changes the answer. Its harder split uses specifiers like before, after, first and last whose dates do not appear directly in the text, which is how people ask about business periods too: last quarter, the quarter before the reorganization, our first full year on the new plan.
Model knowledge does not rescue this. The FreshLLMs authors found that models of every size struggled with fast-changing and false-premise questions, and that adding current search evidence to the prompt improved accuracy substantially. In a company corpus, the evidence passed to the model is the whole game, so the period it covers has to be right.
What does last quarter mean before any filter runs?
A temporal filter is only as good as the period it is given, so resolve relative phrases into explicit date ranges first, and record how you did it. Last quarter depends on the date the question was asked, the company's fiscal calendar and whether the quarter in progress counts. Many companies do not run on calendar quarters; the SEC's description of Form 10-Q, for example, refers to reports for each of the first three fiscal quarters of a company's fiscal year, and fiscal years often start in months other than January.
Documents carry several dates, and only one of them is the right one to filter on for a given question. Pick the field deliberately and store it as its own metadata, not as a date buried in the title.
| Date on a document | What it means | Filter on it when the question asks | Common confusion |
|---|---|---|---|
| Period covered | The business period the content describes | What happened in a quarter | Treated as the publish date |
| Published or approved date | When the document became official | What was reported or decided as of a date | A Q2 memo approved in August is filed under Q3 |
| Effective date range | When a policy or price applied | Which rule applied on a date | Superseded policies with no end date |
| Version or restatement date | When a figure was corrected | The current figure for a past period | Restated and original numbers both indexed as current |
| Ingestion or modified date | When the system last touched the file | Almost never for business questions | A bulk migration makes every file look new |
Where do date filters break at quarter boundaries?
Most wrong-quarter answers in filtered systems come from boundary handling, not from a missing filter. Time zones are the first trap. Qdrant documents that its datetime range condition accepts RFC 3339 values and converts timestamps to UTC for comparison, so a record stamped late on the last evening of a quarter in a Pacific time zone can land in the next quarter unless the period boundaries are written with the company's offset.
Partial dates and rounding are the second. Elasticsearch's range query documentation explains that when a date value is incomplete, missing components are filled with defaults, giving the example that a gt value of 2099-12 with the format yyyy-MM becomes 2099-12-01T23:59:59.999999999Z rather than the end of December. It also documents that date math rounding behaves differently for gt, gte, lt and lte, and that the time_zone parameter does not change the value of now, which is always UTC. Each of these can shift a quarter's edge by hours or days.
Documents that span periods are the third. A year-to-date summary, a trailing-twelve-months report or a two-quarter comparison overlaps the requested quarter without being its controlling source. Index those with a start and an end, and decide per question type whether overlap counts as a match.
Test the edges on purpose rather than hoping the evaluation set happens to include them. For every period in the set, add at least one item whose controlling record sits within a day of the start or end, one whose source system stores local time, and one where a document was approved in the following period. If the filter is correct, these items behave like any other; if it is not, they are the first to fail, and they show which boundary rule is wrong instead of leaving a vague drop in precision to investigate.
What should the time-scoped retrieval and answer evaluation set include?
Each item in the set pairs a question with everything needed to judge both the retrieved evidence and the answer. Write the fields before running any system, so the labels cannot drift toward whatever the retriever returned.
| Field | What to record | Illustrative value |
|---|---|---|
| Question and as-of date | The question as a user would ask it and the date it is asked | How many enterprise renewals slipped last quarter? Asked Oct 18, 2026 |
| Resolved period and basis | Explicit start and end dates and whether fiscal or calendar | May 1 to Jul 31, 2026, fiscal Q2 of a year starting Feb 1 |
| Controlling source | The document and version that governs the answer for that period | Q2 renewal review memo, version 2, restated |
| Required evidence set | Passage ids that must be retrieved to support the answer | Memo v2, section 3 |
| Hard negatives | Same-template documents for adjacent periods, superseded versions and overlapping summaries | Calendar Q3 report, memo v1, fiscal Q3 pipeline export, year-to-date summary |
| Permissions | Who may see the controlling source, and the expected behavior for someone who may not | Revenue leadership only; others get no figure |
| Expected answer | The figure or statement with its period and citation | 14 slipped renewals in fiscal Q2, citing memo v2 (illustrative) |
| Abstain or clarify condition | When the right response is to decline or ask | Ask which quarter if the fiscal calendar is unknown |
Which of five retrieved passages controls the answer? (illustrative)
This example is illustrative: the company, its documents and the counts are made up for this page and reflect no actual business or SourceX record. A sales operations analyst asks on October 18, 2026 how many enterprise renewals slipped last quarter. The company's fiscal year starts on February 1, so the last completed quarter is fiscal Q2, May 1 to July 31, 2026. A calendar reading would instead point at July to September.
The retriever returns five passages. Only one is in the reference evidence set.
| Rank | Passage | Period and version | Label | Why |
|---|---|---|---|---|
| 1 | Renewal report, calendar Q3 | Jul to Sep 2026 | Wrong period | Calendar basis; overlaps fiscal Q2 by one month only |
| 2 | Q2 renewal review memo, v1 | Fiscal Q2, superseded | Wrong version | Restated three weeks later; reports 12 slipped renewals |
| 3 | Q2 renewal review memo, v2 | Fiscal Q2, current | Controlling | Reports 14 slipped renewals after restatement |
| 4 | Year-to-date renewal summary | Feb to Sep 2026 | Overlapping, not controlling | Covers the period but cannot isolate it |
| 5 | Q3 pipeline export | Fiscal Q3, in progress | Wrong period | Quarter not closed on the as-of date |
Which metrics expose the wrong-quarter answer?
Ragas documents an ID-based context precision that divides the number of retrieved context ids found in the reference ids by the total number retrieved; here that is 1 of 5, or 0.2. Its standard context precision rewards placing relevant chunks higher in the ranking, so the controlling memo sitting third behind two confusable passages also costs points. Both metrics only work if wrong-period passages are labeled as not relevant, which is why the hard negatives are written into the item.
At a glance
Which metrics expose the wrong-quarter answer
Add measures that name the temporal failure directly, so the report says what went wrong rather than only that precision was low.
- Period-correct recall at k: share of required evidence for the resolved period retrieved in the top k; 1 of 1 here.
- Wrong-period intrusion rate: share of the top k from a different period; 2 of 5 here, the calendar Q3 report and the fiscal Q3 export.
- Superseded-version rate: share of the top k that is an outdated version of a controlling source; 1 of 5 here.
- Answer period accuracy: the answer states fiscal Q2 with its dates, not just last quarter.
- Citation support: the figure in the answer appears in the cited passage; 14 from memo v2 passes, 12 from memo v1 fails.
- Abstention correctness: for users without permission or for an unclosed quarter, no figure is given.
When is a quarter filter the wrong tool?
Not every question with a date in it wants a single-period filter. A question about the current refund policy wants the version in effect today, which is an effective-date question, not a reporting-period one. A trend question, such as how slipped renewals moved over the last four quarters, needs one controlling passage per quarter, and a retriever that returns four copies of the newest quarter has failed even if each passage is correct.
Some questions have no answer yet. If the analyst had asked about this quarter on October 18, the quarter would still be open, and the right answer explains that the period has not closed and offers the latest available figure with its date. Include items like these in the set, labeled with the behavior you expect, so a system is not rewarded for filling the gap with a stale number.
What can a time-scoped evaluation set not catch?
The set tests whether the system respects the periods you labeled, not whether those labels are right. Have the person who owns each source confirm the controlling version and the fiscal calendar, and re-check items when documents are restated, when the corpus is re-chunked or when metadata fields change. A set built from one company's calendar will not cover the quirks of another, such as 52-53 week fiscal years.
Synthetic templates help with volume but share one weakness: real business records contain the drafts, restatements and mislabeled files that make this hard. Keep a share of items drawn from real document histories, reviewed under the right permissions, so the hard negatives are the ones a production corpus actually contains.
Where do multi-period business records for these evaluations come from?
Time-scoped items need document histories that span several periods with their versions intact, which is why consistency over time matters as much as volume. Buyers commonly want two or more years of consistent records, tens of thousands of them in one system, though a smaller set can be worth it when the work is rare.
Multi-period document histories are sourced by SourceX from U.S. businesses when a team asks, not from stock, and the request-data form is where a retrieval team states scope, volume, language, format, timing and licensing needs.
Frequently asked questions
What are temporal filters in business retrieval?
They are metadata conditions that limit retrieval to documents whose period, effective date or version matches the question, applied before or alongside similarity ranking. They depend on each document carrying the right date fields and on the question's period being resolved into explicit dates. Without both, the filter can silently select the wrong quarter.
Should a wrong-quarter passage count as partly relevant?
For questions about a specific period, no. Label it as a hard negative, because an answer built from it is wrong even when the topic matches. Overlapping documents such as year-to-date summaries can be labeled separately so you can see how often they displace the controlling source.
How do you handle fiscal and calendar quarters in one corpus?
Store the period covered with explicit start and end dates and a basis field, rather than labels like Q2 alone. Resolve each question using the company's fiscal calendar and the as-of date, and include items in the evaluation set where the fiscal and calendar readings disagree.
What should the system do when the requested quarter has not closed?
It should say the period is still open and, where useful, offer the latest available figure with its date and source. Include unclosed-period items in the evaluation set with abstention or clarification as the expected behavior, so a system is not rewarded for answering from an earlier quarter.
Is reranking by recency a substitute for a period filter?
No. Recency boosts favor the newest document, which is the wrong one whenever the question asks about an earlier period. Recency can help for questions about current state, but period questions need an explicit range on the period-covered field.
Sources
- The TimeQA paper defines a time-sensitive question as one with a time specifier whose change alters the answer, with a harder split that requires temporal reasoning over specifiers such as before, after, first and last. Source
- The FreshLLMs paper reports that models struggle on fast-changing and false-premise questions and that adding current search evidence to the prompt improves accuracy substantially. Source
- The SEC describes Form 10-Q as a report filed for each of the first three fiscal quarters of a company's fiscal year. Source
- Qdrant's filtering documentation says the datetime range condition supports RFC 3339 formats and converts timestamps to UTC during comparison. Source
- Elasticsearch's range query documentation explains default values for missing date components, rounding rules for gt, gte, lt and lte, and that the time_zone parameter does not affect now. Source
- Ragas documents context precision as rewarding relevant chunks ranked higher, and an ID-based variant that divides retrieved ids found in the reference ids by all retrieved ids. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.