When the Bill Falls: Reusing the Old Index Defers,Never Avoids, Re-encoding in Complete Embedding-Model Migration
Organizations: Northwestern Polytechnical University Xi’an, China
Abstract
Upgrading the embedding model behind a vector index requires re-encoding the corpus. Recent work reuses the old index through adapters or window re-ranking. Across sixteen model pairs on passage and code corpora, plus eleven on Natural Questions (NQ), we find three results. First, reuse never avoids re-encoding: if migration must complete, every strategy eventually encodes every document; reuse only changes when that work is paid and the quality served meanwhile. Second, adapters should be compared with the old model, not only the new one. Of five adapter variants, only same-family LLM bridges beat the old model, and only by a negligible margin. Third, window re-ranking cost can be priced from the old index alone; pre-registered replays on all 8,841,823 passages confirmed the prediction on all three tested pairs. Workload skew also matters: across 36 further pre-registered Zipf replays, the planned family matched on 32. These findings yield three pre-migration rules: whether reuse is worthwhile, which strategy to use, and how deep a window to take. The depth rule priced every window that reached target quality to within 4 parts in 10^5 of the corpus. On NQ, a pre-committed plan selected the family on 10 of 11 pairs and priced every window exactly. Quality inputs transferred less reliably across corpus scale, so the model gap and window quality loss must be measured on the serving index.
Figures & tables
| Finding | Evidence | Where |
| Reuse defers, never avoids, re-encoding | by construction; spread measured | Fig. 1 |
| Bridges must beat the old model | measured: 5 families 15 pairs | Tab. V |
| Window cost = coverage of served windows | measured: every window replay | Tab. VI |
| Independent-window bound | proven; 1,237 uniform + 2,088 Zipf points; held at 8.8M | Fig. 2 |
| Plain window is the default; unions buy quality at 3.2–5.7 the cost; no selector beats a constant | measured; 5 pre-registered hypotheses; 45 seeded replays | Tab. X |
| Skew shrinks the judged sample | predicted: 36 Zipf replays, family 32 of 36; 4 earlier | § V |
| ID | Nominator | Ranker |
|---|---|---|
| S0 rebuild | old index until switch | , then |
| S1 adapter bridge [ 2 ] | via | |
| S2 window re-rank [ 3 ] | via | |
| S3 mapped window | via | |
| S4 window lexical | BM25 [ 11 ] | |
| S5 window mapped lexical | all three |
| Input | Obtained from | Target encoding needed |
|---|---|---|
| one retrieval pass over | none ; 157–159 s per encoder at | |
| retrieval over | none | |
| retrieval over a target-encoded sample | the sample (25% of to match the full-corpus D1 on 10 of 12 pairs) and the queries | |
| target re-ranking of the judged queries’ source windows | the window documents, at most ( of at ) | |
| Every document vector above is one a completed migration needs: a down payment when D1 says migrate, wasted only when it says no. | ||
| ID | source target | type | gap | CKA | replay winner | ||||||
| P0 | bge-base bge-base | control | – | – | S0/rand | ||||||
| P2 | MiniLM-L6 bge-base | cross-family | S2/d r | ||||||||
| P3 | bge-small bge-base | same family | S2/d r | ||||||||
| P4 | bge-m3 bge-base | same family | S2/d r | ||||||||
| P5 | bge-m3 bge-small | same family | S2/d r | ||||||||
| P9 | MiniLM-L6 bge-small | cross-family | S2/d r |
| Pair | Procr. | affine | l-r 64 | l-r 128 | MLP | best | |
|---|---|---|---|---|---|---|---|
| P2 | |||||||
| P3 | |||||||
| P4 | |||||||
| P5 | |||||||
| P5p | |||||||
| P5q |
| Source | Pairs | gap (docs) | rel. | ||
|---|---|---|---|---|---|
| MiniLM-L6 | P2, P9, P10, P11 | 50 | |||
| bge-base | P14 | 34 | |||
| bge-m3 | P4, P5, P12 | 41 | |||
| bge-small | P3, P13 | 31 |
| Pair | on dev (Recall@10) | on TREC-DL (nDCG@10) |
|---|---|---|
| P5 | ||
| P10 | ||
| P4 | ||
| P12 | ||
| P2 | ||
| P11 |
| planner (static) | replay | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Pair | gap | D1 | D2 | D3 | cost | selected | cost | agree | |
| P0 | rebuild | S0 | – | – | S0/rand | 50 | ✓ | ||
| P2 | migrate | S2 | 25 | S2/d r | 50 | ✓ | |||
| P3 | migrate | S2 | 25 | S2/d r | 50 | ✓ | |||
| P4 | migrate | S2 | 25 | S2/d r | 50 | ✓ | |||
| P5 | migrate | S2 | 25 | S2/d r | 50 | ✓ | |||
| Pair | cost | debt | adv. | winner | verdict | ||
|---|---|---|---|---|---|---|---|
| P2 | never † | S4 | – / ✓ | ||||
| P3 | S2 | ✓ / ✓ | |||||
| P4 | S2 | ✓ / ✓ | |||||
| P5 | S2 | ✓ / ✗ | |||||
| P9 | S4 | ✓ / ✓ | |||||
| P10 | S2 | ✓ / ✓ |
| Hypothesis | Falsified if … measured | Outcome | |
|---|---|---|---|
| H1 | rises as falls | at ; (11 pairs) | fals. |
| H2 | beats majority in leave-one-out regret | regret majority’s; 0.00079 vs 0.00085 | holds |
| H3 | low- /low- cell non-empty | exploratory; P9, P11, but | holds |
| H4 | a probe at of gets within | pair outside; 2 of 10 | fals. |
| H5 | signal order beats rand | S2: 7 better, 6 worse ( ); S2/S4/S5: 8, 20 ( ) | fals. |
| Pair | gap | D3 | docs | failed | ||||
|---|---|---|---|---|---|---|---|---|
| P2f | 1.0M | 25 | 264,274 | 0.2643 | ||||
| 8.8M | 50 | 331,388 | 0.0375 | none | ||||
| P3f | 1.0M | 25 | 261,214 | 0.2612 | ||||
| 8.8M | 25 | 332,874 | 0.0376 | (a), (b) | ||||
| P9f | 1.0M | 25 | 264,274 | 0.2643 | ||||
| 8.8M | 50 | 331,388 | 0.0375 | none |
Text Embedding Models
Text embedding models map words, sentences, or documents to vectors for semantic search, retrieval, clustering, and downstream prediction.
See all 138 papers