Swiss AI Hub v0.305.3

Added

• 🦾 Introduced ResilientOpenAILike for Robust LLM Interactions: A new wrapper for OpenAILike models that
  automatically retries structured prediction calls multiple times. This significantly improves the reliability of
  agents interacting with Large Language Models (LLMs) that may intermittently produce malformed structured output.
• 📄 New Architecture Decision Record (ADR) on Reasoning Model Resilience: Documented the challenges with structured
  output from specific reasoning models (Kimi-K2.6, Qwen3.5) on Infomaniak and detailed the architectural decisions to
  mitigate these issues, including the adoption of plain-text verdicts and retry mechanisms.
• ⚡️ New text_verdict Utility Module: A dedicated module providing helpers for requesting and parsing plain-text
  verdicts from LLMs, which is now used by several guard and detection functions for improved reliability and speed.

Changed

• 🔄 Relevance Guards Now Use Plain-Text Verdicts: The agent_description_guard, few_shot_guard, and
  context_sufficient_guard have been converted from expecting structured JSON output to requesting and parsing simple
  plain-text verdicts (e.g., ALLOW/BLOCK, SUFFICIENT/INSUFFICIENT). This change dramatically increases their
  reliability with reasoning models.
• 🚀 Faster and More Reliable Meta-Question Detection: The detect_meta_question function now uses a streamlined,
  single-token plain-text classification (e.g., NORMAL, META_IDENTITY) with reasoning disabled, reducing latency
  from ~4-5 seconds to sub-second on reasoning models without compromising accuracy.
• 🛡️ Guards Fail Open on Unrecognizable Responses: Relevance guards (agent_description_guard, few_shot_guard,
  context_sufficient_guard) now gracefully degrade by accepting a request if the LLM returns an unparseable verdict,
  preventing agent crashes and improving user experience.
• 💬 Strict System Message Placement for Qwen3.5 Compatibility: The do_answer_meta_question and RAGAgent's
  guard-reject branch now ensure that system messages always lead the message list, resolving a
  400 - System message must be at the beginning error with Qwen3.5.
• ⚙️ Centralized Retry Mechanism for All Structured Predictions: All structured prediction calls are now routed
  through the new ResilientOpenAILike wrapper by default, providing a single point of failure mitigation for all
  agents without requiring individual call-site changes.

Fixed

• 🐛 Resolved Agent Crashes Due to Flaky LLM Structured Output: Addressed a critical issue where agents would crash
  when reasoning models like Kimi-K2.6 or Qwen3.5 failed to produce reliably parseable structured output (e.g., empty
  tool calls, malformed JSON).
• 🔒 Prevented Sensitive Information Leaks in Guard Logs: Guard-failure logs no longer include raw model output,
  queries, or retrieved context, enhancing privacy and security.