search-reranker-broad-policy-v3

search-reranker-broad-policy-v3

Broad policy reranker with stronger office-holder behavior and a better shipped q8 profile.

This release keeps the structured candidate format from v1 / v2:

• optional Category: ...
• optional Published: YYYY-MM-DD
• multi-turn queries passed as newline-separated text

The main change is the deployed path:

• small weight continuation from v2
• dynamic-int8 ONNX with reduce_range
• still use office_holder_v1 for role/person identity queries

Recommended serving profile

• backend: onnx
• artifact: onnx/model_quantized.onnx
• threads: 8
• max length: 224
• policy: office_holder_v1 for identity queries, none otherwise

Key q8 results

Why this release matters

• It improves the broad policy suite on the shipped q8 path.
• It keeps the fresh policy holdout stable.
• It materially improves office-holder identity ranking on the shipped q8 path.
• It is faster than the previous broad-policy-v2 q8 office-policy path.

Runtime

Office-policy validation set, q8, threads=8, max_length=224, office_holder_v1:

• broad-policy-v2: 13.05 qps
• broad-policy-v3: 23.91 qps

Policy holdout v3, q8, threads=8, max_length=224:

• broad-policy-v3: 28.86 qps

Notes

• The fp32 ONNX export preserves the small weight-level gains more faithfully than standard qint8 per-channel quantization.
• The included q8 artifact therefore uses reduce_range, because it retained the gains while keeping CPU speed in the same class.
• If Category: / Published: are absent, the reranker still works; they are optional metadata cues, not required features.

Portfolio comparison

<!-- portfolio-comparison:start -->

Updated 2026-03-20 from local reranker reports only.

Use this section as the side-by-side public temsa reranker view. Cells are intentionally left out or summarized when the local report set does not contain a trustworthy non-quantized benchmark for the same public path.

General bilingual rerankers

Policy rerankers

Dynamic q8 CPU at K=20

These rows use the current local K=20 benchmark harness:

• proxy_k20: policy=none
• broad_policy_k20: policy=gov_broad_v1
• office_holder_k20: policy=office_holder_v1
• fixed threads=10, batch_size=32

Intentional gaps:

• search-reranker-broad-v1: the local reports include strong fp32 ONNX proxy and office data, but not a matching fp32 in-domain finephrase / hard-K10 / holdout-A03 set, so those are not claimed here.
• search-reranker-irishgov-l5-k10-v2: the local non-quantized reports are trustworthy for office / finephrase / hard-K10, but not for the same proxy runtime shape as the shipped q8 path.
• search-reranker-irishgov-l6-fast-v1: the local non-quantized reports cover proxy and office, but not the fresh holdout-A03 slice used for the q8 K=10 comparison.

<!-- portfolio-comparison:end -->