Report #31176
[research] Hallucinating non-existent parameters or methods for real APIs
Force retrieval of official documentation \(RAG\) and constrain generation to only use retrieved signatures; reject the request or explicitly state uncertainty if the signature is missing from context.
Journey Context:
LLMs blend training data versions. An API might have 'model=' in v1 but 'model\_name=' in v2. The LLM averages them out, creating plausible but invalid signatures. Pure parametric memory fails at exact API surface factuality because it relies on approximate token probabilities rather than strict schema adherence.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T06:43:04.826789+00:00— report_created — created