Implement AI solutions by using Microsoft Foundry

System prompts are the primary mechanism for defining agent behavior, role boundaries, and response patterns. Including explicit instructions about what topics to avoid and how to redirect users is the most effective and reliable method. System prompts can provide specific examples like 'If asked about competitors, respond: I'm here to help with [Product] questions. For other products, please contact their support teams.'

Azure AI Language uses native language-specific models for key phrase extraction. Different languages have different grammatical structures, idiomatic expressions, and cultural contexts. A concept expressed in 2 words in English might require 4 words in French or be captured in a compound word in German. The models are trained to identify what constitutes a 'key phrase' in each language's natural structure, so output length variation is expected and correct behavior.

The prebuilt invoice model is trained on thousands of invoice variations and can handle most vendor layouts out-of-the-box. It automatically extracts common fields (invoice ID, dates, amounts, line items, tables) and handles handwriting. This is the most cost-effective option - no training data required, no training costs, immediate deployment. For the small percentage of fields it might miss due to unusual layouts, targeted post-processing is cheaper than training custom models.

Object detection is the correct capability because users will photograph plates with MULTIPLE food items simultaneously. Object detection identifies each item, draws bounding boxes around them, and classifies each one separately (e.g., one image might contain apple, rice, AND chicken). This is exactly what nutritional estimation needs - identify each distinct food item present. Custom training is needed because food recognition is domain-specific.

Speech translation (part of Azure AI Speech) performs speech recognition and translation in a SINGLE operation with auto-detect. It outputs translated text directly without needing a separate transcription→translation pipeline. This is the lowest latency option because it eliminates the round-trip to a separate Translator service call and processes both operations in the speech pipeline.

Facial DETECTION (detecting that a face exists and its location) is different from facial RECOGNITION/identification. Detection only returns rectangle coordinates and attributes - it doesn't create biometric templates or identify individuals. This can be used in HIPAA environments if the overall application meets compliance requirements (encryption, access controls, BAA with Microsoft, etc.). The detection itself doesn't create restricted biometric data.