We build AI agents, voice agents & business automation for growing teams. CRM integrations, RAG chatbots & web apps. Tell us what you want to automate.
AGI Innovations builds AI & voice agents, CRM automation, RAG chatbots and custom web apps for growing businesses. Have a project? Tell us your workflow, current tools and goal: https://t.co/PKPIXQ11rL
Before shipping a support chatbot, give it two versions of the same refund policy. Ask which rule applies today and check the cited passage. Then remove the current version and verify that it escalates instead of guessing. #RAG#AIAutomation
@mchulet For the RAG portfolio project, show a small eval set with answerable questions, missing answers and superseded docs. Report where it abstains and whether the cited passage supports the answer. That’s easier to assess than the document count.
@harikesh_devops@Qasimshah710 A useful freshness test: keep an archived runbook and its replacement in the index. Ask what to do today, then ask what the process was last month. The answer should cite the version that was effective at the time.
@contextaware_ai I’d test a pause inside a phone number. If the caller says the first few digits, stops to check the rest, then continues, the agent should wait without treating the pause as the end of the turn. Background noise makes this worth testing on a real call.
@BoboLinNotes For a document extraction portfolio project, I’d include a review queue and show how it handles unreadable scans or conflicting totals. A short demo of the operator resolving one exception tells me more about readiness than a perfect sample PDF.
@imaskie A useful first check for parallel-only flakes is to run two workers against the same test account. Shared fixtures can make one test delete or modify another’s data. Give each worker isolated test data before adding more retries.
@JakeDSBD24 For that end-to-end test, carry the same job ID from the 202 response through to the docs update. Assert the terminal state and check that the generated docs reference the expected commit. That catches jobs that get accepted and then disappear.
@zahragr8r That copy gives the user something they can actually fix. I’d add a “Send test message” button next to the connection setting so they can verify the fix without rerunning the entire automation.
@WaqasSethi The second call is a great test. I’d have the agent resume from the last unresolved issue, then test a caller who says the situation changed. The new information should take precedence and update the record for the next handoff.
@telnyx I’d use caller identity to find a candidate record and step up verification before reading private details or changing an account. A good test is two customers sharing a number. The greeting can stay useful without assuming who picked up.
@jlazer78 Linking the highlighted pins back to the original photo is a useful inspection detail. How do you handle a finding where the scan has a blind spot? Showing the view that supports each finding would help an operator decide when to rescan.