RetrieX Numeric Extreme Retrieval Fix

Purpose

This patch sharpens retrieval for direct numeric extreme questions such as the lowest hardness threshold.

The concrete regression was:

User asks for the lowest water-hardness threshold monitored by a Testomat.
The correct answer is 0,02 °dH / Testomat 808.
Retrieval still allowed neighbouring runner-up product context such as Testomat 2000 / 0,05 °dH into the prompt.

That made the model add unnecessary comparison details although the user asked only for the lowest value.

Change

src/Knowledge/Retrieval/NdjsonHybridRetriever.php now adds a conservative numeric-extreme document selection step between focused-product selection and normal dominant/spread selection.

The new mode:

detects minimum/maximum-style technical measurement questions,
extracts dH measurement values from the top retrieval window,
identifies the document containing the actual extreme value,
selects chunks from that document only,
avoids filling the remaining prompt slots with runner-up product chunks.

New debug selection mode:

sales_numeric_extreme_document

Safety

The fix is intentionally narrow:

no PromptBuilder changes,
no prompt wording changes,
no Shopware logic changes,
no vector-service changes,
no scoring config changes,
no vocabulary changes.

It only affects technical numeric extreme questions containing measurement/context signals such as Grenzwert, Messbereich, Wasserhärte, Resthärte, dH, threshold, or range.

Expected regression result

Question:

Was ist der niedrigste Grenzwert für die Wasserhärte, welcher mit einem Testomaten überwacht werden kann?

Expected answer should stay focused on:

0,02 °dH / Testomat 808

It should not add the runner-up product/value such as:

Testomat 2000 / 0,05 °dH

unless the user explicitly asks for comparison, alternatives, or all available values.

After applying

Run:

php bin/console cache:clear
php bin/console mto:agent:config:validate
php bin/console mto:agent:regression:test

Then manually retest the known 1.4.2 baseline and the lowest-threshold prompt above.

2.2 KiB Raw Blame History