patch 20o

config/retriex/agent.yaml

@@ -24,10 +24,10 @@ parameters:
         intro: 'Normalize the following user input for RetrieX routing before intent detection.'
         output_format_block: |-
           Output format:
-          <normalized user input>
+          Return exactly one line: the corrected user input.
         current_user_input_label: 'USER INPUT'
         rules:
-          - '- Output only the normalized user input.'
+          - '- Output only the corrected user input text itself, never the words "normalized user input".'
           - '- Correct only obvious typing mistakes, transposed letters, missing umlauts, spacing, and punctuation that clearly preserve the same meaning.'
           - '- Do not add product names, model numbers, article numbers, measurement values, parameters, brands, or application areas that are not already present in the input.'
           - '- Preserve product names, model numbers, article numbers, chemical symbols, units, pH, Redox, ORP, and measurement values exactly unless only letter casing is corrected.'
@@ -410,7 +410,7 @@ parameters:
         - '- Preserve the language of the CURRENT USER INPUT for generic product/search terms; do not translate German search terms into English.'
         - '- For German user input, output German shop terms, for example "freies Chlor Messung" instead of "free chlorine measurement".'
         - '- Preserve domain terms from the current user input or resolved context in their original language.'
-        - '- Numbers that belong to a product name or model must be preserved (e.g. Indikator 300, Testomat 808, Testomat 2000).'
+        - '- Numbers that belong to a product name or model must be preserved when they are present in the CURRENT USER INPUT or a clearly resolved reference.'
         - '- Separate terms using spaces only.'
         - '- If a relevant product name is present, it must be placed at the beginning of the final search query.'
         - '- Try to always identify all products mentioned in the user input text, even in long prompts.'