Can Vectara find new content after it is added to the source?
Vectara made a newly added document findable right away: before the add no retrieved passage mentioned the Tromso depot, and afterwards all four identical queries retrieved the passage giving its opening date. Only two of those four answers stated the date; the other two declined.
1 of 1 test case passed
Every test case in this scenario has a result.
The test case
Each test case is judged on its own: Pass, Fail, Not gradable, or Untested. The scenario result above is based on this test case.
Document added to the source, then queried after a refreshPassEvidence
“When does the Tromso depot open?”
Query 1 of 4: date given, passage at rank 1 The Tromso depot opens on 14 October 2026 [1]. Query 2 of 4: declined, passage at rank 8 I do not have enough information to determine when the Tromso depot opens. Query 3 of 4: date given, passage at rank 3 The Tromso depot opens on 14 October 2026 [3]. Query 4 of 4: declined, passage at rank 6 I do not have enough information to determine when the Tromso depot opens.
Transcribed from Proof 1 · Screenshot
| ✓ | Expected: The added fact comes back after the update; how long the update takes is recorded but not graded.Found: Before the add, no passage mentioned Tromso. After it, all four queries retrieved the passage with 14 October 2026, at ranks 1, 8, 3 and 6; two answers gave the date. |
TC98 — S88 · New content is added to the source — full evidence ====================================================================== Capability: C36 Knowledge Sync · Query: 'When does the Tromso depot open?' Reference answer: 14 October 2026 Mutation: MS-v0 -> MS-v1 (add B05_tromso_depot_bulletin.md) Corpus: r6-managed-rag-ms-v1 (separate corpus from the SS-v1 stable slice) Slice digests recorded (R6_slice_versions.json): MS-v0: 0c40b34f1fb09799b512bb937465646d14e0a0b2fbf94a6b816ceb4e9d35ec1c MS-v1: 953bb6960d50a9697757a6161813f96cb4b6044e79161df4a585d4cd02907b0a ---------------------------------------------------------------------- STEP 1 — MS-v0 baseline ingest (4 files) ---------------------------------------------------------------------- Documents present: 4 - B01_oncall_rota_notes.md - B02_shipping_lead_times.pdf - B03_approved_supplier_list.docx - B04_site_access_notes.txt ---------------------------------------------------------------------- STEP 2 — pre-add query against MS-v0 (baseline, before the mutation) ---------------------------------------------------------------------- Generated answer: I do not have enough information to determine when the Tromso depot opens. Top 5 retrieved passages: rank 1 | score=0.8058 | doc=B02_shipping_lead_times.pdf Lead times to the Nordic depots assume a groupage service. rank 2 | score=0.7607 | doc=B02_shipping_lead_times.pdf Central store, Bristol - outbound to depots and customer sites rank 3 | score=0.7504 | doc=B02_shipping_lead_times.pdf Anything handed over after the 15:30 cut-off travels the following working day, and the lead time starts then. rank 4 | score=0.6125 | doc=B04_site_access_notes.txt Codes are held by the service desk and read out on the day. rank 5 | score=0.6068 | doc=B03_approved_supplier_list.docx Review The stores controller reviews this list each quarter with the workshop manager. Any of the 10 retrieved passages mentions Tromso: False This confirms the fact is not inferable/present anywhere in MS-v0 before the mutation. ---------------------------------------------------------------------- STEP 3 — apply the mutation: add B05_tromso_depot_bulletin.md ---------------------------------------------------------------------- Upload response: HTTP 201 (add started 2026-09-10T21:53:58Z, response 21:54:02Z — 4s) Documents present after the add: 5 - B01_oncall_rota_notes.md - B02_shipping_lead_times.pdf - B03_approved_supplier_list.docx - B04_site_access_notes.txt - B05_tromso_depot_bulletin.md ---------------------------------------------------------------------- STEP 4 — post-add queries: 4 independent, identical queries against the same post-add corpus state (no further mutation between them) ---------------------------------------------------------------------- run A (first query, immediately after the add) answering chunk rank: 1/10 generated answer: 'The Tromso depot opens on 14 October 2026 [1].' NOTE: this run's raw JSON was captured and printed live, but was inadvertently overwritten on disk by a later duplicate call before being archived under its own filename. The values above are the actual response as printed and observed at the time. top passages (from the same live printout): rank1 B05 'The Tromso depot opens on 14 October 2026...' (0.7874), rank2 B05 'For the service desk, the Bergen depot...' (0.7445), rank3 B05 'Depot bulletin - a new depot...' (0.7291) run B (duplicate re-run, captured on disk) answering chunk rank: 8/10 generated answer: 'I do not have enough information to determine when the Tromso depot opens.' factual_consistency_score: 0.22460938 full retrieved ranking: 1. score=0.9312 'Depot bulletin - a new depot in the region' 2. score=0.8858 'It is the third depot in the network and the firstabove the Arctic Cir' 3. score=0.7965 'The new depot does not calibrate, does notbreak a calibration seal, an' 4. score=0.7965 'Stop routing northern loan requests through Bergen once the depot open' 5. score=0.7965 'Out of hours the northern sites continue to page the existingduty engi' 6. score=0.7937 'Ringthe new depot directly.' 7. score=0.5511 'For the service desk, the Bergen depot and central stores.' 8. score=0.4808 'The Tromso depot opens on 14 October 2026 and will hold four Tolvan C5' 9. score=0.3512 'One engineer and a shared stores controller, working 07:00 to 15:00 lo' 10. score=0.0989 'What is opening' run C (captured on disk) answering chunk rank: 3/10 generated answer: 'The Tromso depot opens on 14 October 2026 [3].' factual_consistency_score: 0.46289062 full retrieved ranking: 1. score=0.9428 'Depot bulletin - a new depot in the region' 2. score=0.9364 'The new depot does not calibrate, does notbreak a calibration seal, an' 3. score=0.8782 'The Tromso depot opens on 14 October 2026 and will hold four Tolvan C5' 4. score=0.8336 'Out of hours the northern sites continue to page the existingduty engi' 5. score=0.7953 'Ringthe new depot directly.' 6. score=0.7302 'What is opening' 7. score=0.4958 'One engineer and a shared stores controller, working 07:00 to 15:00 lo' 8. score=0.4912 'For the service desk, the Bergen depot and central stores.' 9. score=0.1280 'It is the third depot in the network and the firstabove the Arctic Cir' 10. score=0.1204 'Stop routing northern loan requests through Bergen once the depot open' run D (captured on disk) answering chunk rank: 6/10 generated answer: 'I do not have enough information to determine when the Tromso depot opens.' factual_consistency_score: 0.3125 full retrieved ranking: 1. score=0.9401 'One engineer and a shared stores controller, working 07:00 to 15:00 lo' 2. score=0.9062 'For the service desk, the Bergen depot and central stores.' 3. score=0.7640 'It is the third depot in the network and the firstabove the Arctic Cir' 4. score=0.5911 'What is opening' 5. score=0.3571 'Depot bulletin - a new depot in the region' 6. score=0.1657 'The Tromso depot opens on 14 October 2026 and will hold four Tolvan C5' 7. score=0.1657 'The new depot does not calibrate, does notbreak a calibration seal, an' 8. score=0.1657 'Stop routing northern loan requests through Bergen once the depot open' 9. score=0.0104 'Out of hours the northern sites continue to page the existingduty engi' 10. score=0.0104 'Ringthe new depot directly.' ---------------------------------------------------------------------- FINDING — retrieval always finds the fact; ranking of the specific chunk is unstable ---------------------------------------------------------------------- Across all 4 identical post-add queries (A, B, C, D) against the unchanged 5-document corpus, the added B05 document is always present, and the exact chunk stating '14 October 2026' always appears somewhere in the top 10 search_results (ranks: 1, 8, 3, 6). The fact never became unfindable/omitted at any point after the add — no indexing delay was observed (the very first query after the 201 upload response already surfaced it at rank 1). However, the *rank* of that chunk swings widely run-to-run with no further state change, which in 2 of 4 runs pushed it low enough that generation declined instead of answering (runs B and D), while in the other 2 it was ranked high enough to be cited correctly (runs A and C). This is a generation-layer/reranker-instability finding, not a knowledge- sync failure: TC98's own grading surface is retrieval ('the fact comes back'), and by that criterion it passed in all 4 runs. It is documented here because it directly affects whether an end user asking this question through Vectara's generation layer gets a usable answer on any single try. ---------------------------------------------------------------------- UI verification (console.vectara.com) ---------------------------------------------------------------------- Navigated to the mutation corpus (r6-managed-rag-ms-v1) Data tab live in the browser. The console shows '5 documents' and lists B05_tromso_depot_bulletin.md alongside the 4 MS-v0 baseline files — confirmed independently of the API responses above. Screenshot: tc98-vectara-console-mutation-corpus-documents.png
This test fails only if the added fact never becomes reachable. The passage with the date was retrieved on every query after the add, so the result is a Pass; how often the written answer stated the date is not what this test grades.
Configuration and setup
How this tool was set up for the run and what the test needed in place. Each row is a fact from the run's records; a fact the records do not hold is left out, not guessed.
How this scenario is graded
How we decide Pass, Fail and Not gradable. The same rules apply to every tool tested on this scenario.
Each test case gets one result per tool: Pass, Fail or Not gradable. A test case we haven't run yet shows Untested. There are no partial results.
- Pass: the tool did everything the test expected, and nothing it said contradicts the correct answer.
- Fail: at least one expected behaviour clearly didn't happen; the test case's result says which one and quotes the tool.
- Not gradable: our evidence couldn't settle the outcome (for example a record we needed is missing). It is never counted as a fail, and the test case's result says what's missing.
Where this sits in the benchmark
This page is one cell of a larger study: one tool, one scenario. Only this benchmark's frame appears here.
| Level | Name | Scope |
|---|---|---|
| Benchmark | Managed RAG → | 15 scenarios · 9 tools |
| Capability | Knowledge Sync → | |
| Scenario | New content is added to the source → | |
| Tool | Vectara → |
Global scenario definition → · Global capability definition → · Vectara product page →
History of this result
What has happened to this result since it was first published. Runs and grades are never overwritten: a retest or a re-grade publishes a new result and keeps the earlier one readable.
Act on this result
Nothing filed here edits the run or the grade. A challenge opens a review, and a review can produce a new run or a re-grade — which becomes the current result and leaves this one in the history.
You run the same kind of test against your own setup and get the same behaviour.
Agree →Yours behaves differently. Tell us what you got, with a screenshot if you have one.
Disagree →Something here is wrong — a reference value, a transcription, a grade.
Report an issue →Tell us what changed and we schedule a rerun of the failing test case. The old result stays as history.
Vendor notice →The same record is available as structured data through the AI Demos MCP server, with the counts, the coverage and every per-test-case reason carried as fields.
Verify the proof files
These files support this result. Open a file to inspect the original evidence.
File fingerprints (SHA-256)
A fingerprint identifies the exact file used for this result.