Day 32 of building DocuPulse AI.

Today: if OpenAI goes down, the system keeps answering by falling back to a second provider. Sounds like a one-line feature. Testing it taught me two things I didn't expect.

First — I broke my OpenAI key to simulate an outage, and the system died before the failover ever kicked in. Why? The failover was on the answer step, but the retrieval step — turning the question into a vector to search my documents — also uses OpenAI, and it can't fall back. My stored search index is in OpenAI's "vector space"; another provider's vectors simply wouldn't match it. So failover can keep the system answering, but it can't keep it searching. Real limitation, worth knowing.

Second — when the backup provider did take over, it gave a correct answer, but formatted completely differently. More verbose, extra structure. Even when I tested a question with no answer in the document, both providers correctly refused to guess — but one said exactly "I do not know" and the other said "I do not know" plus a paragraph explaining why. If my code checked for that exact phrase, the failover would silently break it.

The lesson: failover doesn't just change whether you get an answer. It changes who answers and how they behave. You have to test the backup path as carefully as the main one — a different model is a different system, not a drop-in clone.

#AIEngineering #Laravel #BackendEngineering #BuildInPublic