p99c
This commit is contained in:
@@ -1286,6 +1286,13 @@ parameters:
|
||||
- schwimmbad
|
||||
- schwimmbecken
|
||||
- pool
|
||||
- silikat
|
||||
- silikatüberwachung
|
||||
- silikatueberwachung
|
||||
- sio2
|
||||
- si o2
|
||||
- kieselsäure
|
||||
- kieselsaeure
|
||||
- 0,02
|
||||
stopword_cleanup:
|
||||
origin: genre_native
|
||||
|
||||
@@ -0,0 +1,85 @@
|
||||
# RetrieX Patch p99b - Eval Suite Alignment
|
||||
|
||||
## Ziel
|
||||
|
||||
p99 hatte die neue Eval-Suite erfolgreich aktiviert, aber drei neue Cases zeigten nach dem ersten Lauf rote Signale. p99b trennt dabei False-Positive-Assertions von zwei realen Robustheitsluecken, ohne die bestehende Retrieval-Baseline oder Shop-/Follow-up-Architektur umzubauen.
|
||||
|
||||
## Ausgangslage
|
||||
|
||||
Nach p99:
|
||||
|
||||
- `mto:agent:config:validate`: OK
|
||||
- `mto:agent:eval:run retrieval`: 19/19 OK
|
||||
- `mto:agent:eval:run shop_query`: 4/5 OK
|
||||
- `mto:agent:eval:run followup`: 3/4 OK
|
||||
- `mto:agent:eval:run answer_guard`: 3/4 OK
|
||||
|
||||
Rote Cases:
|
||||
|
||||
- `shop_query_sio2_anchor_001`: normalisierte Shopquery konnte auf `gerät` zusammenschrumpfen.
|
||||
- `followup_main_device_price_001`: Hauptgeraet-Follow-up konnte an der vorherigen Indikator-Query `testomat 808 indikator 300` haengen bleiben.
|
||||
- `answer_guard_delivery_not_sdb_001`: Assertion war zu streng, weil ein Textbegriff `Sicherheitsdatenblatt` im Retrieval-Text kein ausreichender Fehlernachweis ist, solange das falsche Dokument nicht dominiert.
|
||||
|
||||
## Aenderungen
|
||||
|
||||
### 1. SiO2/Silikat als aktuelle Eingabe schuetzen
|
||||
|
||||
`config/retriex/genre.yaml`
|
||||
|
||||
Ergaenzt `shop_query_runtime.current_input_preservation_terms` um:
|
||||
|
||||
- `silikat`
|
||||
- `silikatüberwachung`
|
||||
- `silikatueberwachung`
|
||||
- `sio2`
|
||||
- `si o2`
|
||||
- `kieselsäure`
|
||||
- `kieselsaeure`
|
||||
|
||||
Damit verliert eine normalisierte Standalone-Shopfrage wie `suche gerät kühlsysteme Silikatüberwachung` nicht mehr den fachlichen Messparameter, bevor die generische Device-Anchor-Regel `testomat 808 sio2` greifen kann.
|
||||
|
||||
### 2. Hauptgeraet-Follow-up darf Zubehoerreste entfernen
|
||||
|
||||
`src/Agent/AgentRunner.php`
|
||||
|
||||
`guardMainDeviceReferentialShopQueryWithHistoryModelAnchor()` wurde so angepasst, dass eine Shopquery wie `testomat 808 indikator 300` bei einem Prompt wie `und was kostet das gerät selber` nicht allein deshalb akzeptiert wird, weil sie bereits einen Modellanker enthaelt.
|
||||
|
||||
Neu wird geprueft, ob nach dem Modellanker noch Zubehoer-/Code-Resttokens vorhanden sind. Falls ja, wird auf den reinen Modellanker aus dem Verlauf reduziert, z. B. `testomat 808`.
|
||||
|
||||
### 3. Answer-Guard-Case weniger spröde
|
||||
|
||||
`tests/evals/cases/answer_guard.ndjson`
|
||||
|
||||
Der Case `answer_guard_delivery_not_sdb_001` prueft weiterhin:
|
||||
|
||||
- passendes Liefer-/Versand-Dokument muss enthalten sein
|
||||
- konkretes SDB-Dokument darf nicht enthalten sein
|
||||
|
||||
Die zu breite Text-Assertion auf den Begriff `sicherheitsdatenblatt` wurde entfernt, weil sie auch legitime Neben-/Hinweistexte treffen kann.
|
||||
|
||||
## Bewusst nicht geaendert
|
||||
|
||||
- Keine Retrieval-Gewichte
|
||||
- Keine Shopware-Suche
|
||||
- Keine Prompt-Texte
|
||||
- Keine Modellparameter
|
||||
- Keine neue Produkt-Sonderlogik
|
||||
- Keine Aenderung an p98-Retrieval-Eval-Cases
|
||||
|
||||
## Erwartete Checks
|
||||
|
||||
```bash
|
||||
php bin/console mto:agent:config:validate
|
||||
php bin/console mto:agent:eval:run retrieval
|
||||
php bin/console mto:agent:eval:run shop_query
|
||||
php bin/console mto:agent:eval:run followup
|
||||
php bin/console mto:agent:eval:run answer_guard
|
||||
```
|
||||
|
||||
Erwartung:
|
||||
|
||||
- Config valid
|
||||
- Retrieval 19/19
|
||||
- Shopquery 5/5
|
||||
- Followup 4/4
|
||||
- Answer guard 4/4
|
||||
@@ -0,0 +1,60 @@
|
||||
# RETRIEX PATCH 99C - Main Device Follow-up Eval Alignment
|
||||
|
||||
Status: patch-only follow-up for p99/p99b.
|
||||
|
||||
## Goal
|
||||
|
||||
Keep the new p99 follow-up eval suite aligned with the already confirmed manual
|
||||
reference flow:
|
||||
|
||||
1. lowest water-hardness threshold
|
||||
2. indicator type
|
||||
3. indicator price
|
||||
4. main device price
|
||||
|
||||
The main-device follow-up `und was kostet das gerät selber` must resolve back to
|
||||
the main device anchor (`testomat 808`) and must not keep accessory remnants such
|
||||
as `indikator` or exact indicator code `300`.
|
||||
|
||||
## Root cause
|
||||
|
||||
p99b added a residual accessory guard, but the main-device history-anchor guard
|
||||
returned early for non-generic shop queries before the residual check could run.
|
||||
A query like `testomat 808 indikator 300` contains digits, so it was not treated
|
||||
as a generic main-device query and stayed unchanged.
|
||||
|
||||
## Change
|
||||
|
||||
`AgentRunner::guardMainDeviceReferentialShopQueryWithHistoryModelAnchor()` now:
|
||||
|
||||
1. detects the main-device referential prompt,
|
||||
2. extracts the latest history model anchor,
|
||||
3. if the generated shop query already contains that model anchor, checks for
|
||||
accessory/code residuals,
|
||||
4. reduces the query to the pure model anchor when such residuals are present.
|
||||
|
||||
This keeps explicit non-generic product queries untouched unless they contain the
|
||||
current history model anchor plus accessory leftovers in a main-device follow-up.
|
||||
|
||||
## Expected eval result
|
||||
|
||||
```bash
|
||||
php bin/console mto:agent:config:validate
|
||||
php bin/console mto:agent:eval:run retrieval
|
||||
php bin/console mto:agent:eval:run shop_query
|
||||
php bin/console mto:agent:eval:run followup
|
||||
php bin/console mto:agent:eval:run answer_guard
|
||||
```
|
||||
|
||||
Expected:
|
||||
|
||||
- retrieval: 19/19
|
||||
- shop_query: 5/5
|
||||
- followup: 4/4
|
||||
- answer_guard: 4/4
|
||||
|
||||
## Productive logic impact
|
||||
|
||||
Minimal. The patch only changes the already existing main-device follow-up guard
|
||||
for prompts asking for the main device itself. It does not modify retrieval,
|
||||
ranking, prompt templates, YAML vocabulary, shop result guards, or answer logic.
|
||||
@@ -4155,7 +4155,6 @@ final readonly class AgentRunner
|
||||
$shopSearchQuery === ''
|
||||
|| trim($commerceHistoryContext) === ''
|
||||
|| $this->referenceAnchorExtractor->extractFirstProductModelAnchor($prompt) !== ''
|
||||
|| $this->referenceAnchorExtractor->extractFirstProductModelAnchor($shopSearchQuery) !== ''
|
||||
) {
|
||||
return $shopSearchQuery;
|
||||
}
|
||||
@@ -4164,10 +4163,6 @@ final readonly class AgentRunner
|
||||
return $shopSearchQuery;
|
||||
}
|
||||
|
||||
if (!$this->isGenericMainDeviceReferentialShopQuery($shopSearchQuery)) {
|
||||
return $shopSearchQuery;
|
||||
}
|
||||
|
||||
$modelAnchor = $this->normalizeShopQueryAnchor(
|
||||
$this->extractLatestHistoryProductModelAnchor($commerceHistoryContext)
|
||||
);
|
||||
@@ -4176,9 +4171,43 @@ final readonly class AgentRunner
|
||||
return $shopSearchQuery;
|
||||
}
|
||||
|
||||
return $this->queryAlreadyContainsAllAnchorTokens($shopSearchQuery, $modelAnchor)
|
||||
? $shopSearchQuery
|
||||
: $modelAnchor;
|
||||
if ($this->queryAlreadyContainsAllAnchorTokens($shopSearchQuery, $modelAnchor)) {
|
||||
return $this->containsMainDeviceFollowUpAccessoryResidual($shopSearchQuery, $modelAnchor)
|
||||
? $modelAnchor
|
||||
: $shopSearchQuery;
|
||||
}
|
||||
|
||||
if (!$this->isGenericMainDeviceReferentialShopQuery($shopSearchQuery)) {
|
||||
return $shopSearchQuery;
|
||||
}
|
||||
|
||||
return $modelAnchor;
|
||||
}
|
||||
|
||||
private function containsMainDeviceFollowUpAccessoryResidual(string $shopSearchQuery, string $modelAnchor): bool
|
||||
{
|
||||
$queryTokens = $this->tokenizeShopQueryCandidate($shopSearchQuery);
|
||||
if ($queryTokens === []) {
|
||||
return false;
|
||||
}
|
||||
|
||||
$modelTokens = array_fill_keys($this->tokenizeShopQueryCandidate($modelAnchor), true);
|
||||
$accessoryTokens = $this->buildShopQueryTokenSet($this->mergeUniqueStrings(
|
||||
$this->agentRunnerConfig->getNoLlmAccessoryProductRoleKeywords(),
|
||||
$this->agentRunnerConfig->getRequestedAccessoryCodeTerms()
|
||||
));
|
||||
|
||||
foreach ($queryTokens as $token) {
|
||||
if (isset($modelTokens[$token])) {
|
||||
continue;
|
||||
}
|
||||
|
||||
if (isset($accessoryTokens[$token]) || preg_match('/^\d{1,5}$/u', $token) === 1) {
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
private function guardWeakReferentialShopQueryWithHistoryModelAnchor(
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
{"id":"answer_guard_noise_no_evidence_001","type":"answer_guard","prompt":"dsgfsdgfsdgf","assert":{"max_results":0}}
|
||||
{"id":"answer_guard_mythical_medium_no_direct_evidence_001","type":"answer_guard","prompt":"gibt es einen testomat für drachenblut","assert":{"must_not_include_terms":["drachenblut"]}}
|
||||
{"id":"answer_guard_lunar_water_no_direct_evidence_001","type":"answer_guard","prompt":"welcher testomat misst mondwasser im vakuum","assert":{"must_not_include_terms":["mondwasser","vakuum"]}}
|
||||
{"id":"answer_guard_delivery_not_sdb_001","type":"answer_guard","prompt":"lieferbedingungen versand testomat","assert":{"min_results":1,"must_include_one_of_document_ids":["26ddf03d-9108-4a65-aa0e-a5df7613fa77"],"must_not_include_document_ids":["7166592f-85f2-425c-997b-73e323ae184d"],"must_not_include_terms":["sicherheitsdatenblatt"]}}
|
||||
{"id":"answer_guard_delivery_not_sdb_001","type":"answer_guard","prompt":"lieferbedingungen versand testomat","assert":{"min_results":1,"must_include_one_of_document_ids":["26ddf03d-9108-4a65-aa0e-a5df7613fa77"],"must_not_include_document_ids":["7166592f-85f2-425c-997b-73e323ae184d"]}}
|
||||
|
||||
Reference in New Issue
Block a user