AI Troubleshooting

When working with intelligent systems, unexpected behaviors may occur — from irrelevant results to model refusals or latency issues. This section helps you identify, understand, and resolve common AI-related problems in Neolens.

⚠️ Common Issues

🔄 Inconsistent Outputs

• Symptoms: Same input yields different results.
• Causes: Model stochasticity, non-fixed temperature value.
• Fixes:
  • Set temperature to 0.0 for deterministic output.
  • Use a fixed seed (if available).

🙈 Irrelevant or Incomplete Analysis

• Symptoms: Output is vague, unrelated, or missing key anatomical data.
• Causes: Ambiguous input, unrecognized data format, lack of signal in image.
• Fixes:
  • Double-check input file format and quality.
  • Validate that anatomical region is supported.
  • Try adjusting the prompt or re-uploading the image.

🛑 AI Refuses to Generate Output

• Symptoms: Model refuses, citing uncertainty or safety concern.
• Causes: Image too ambiguous or sensitive content detection triggered.
• Fixes:
  • Provide clearer inputs.
  • Lower the safety threshold (if adjustable).
  • Check logs for flagged content.

---

💡 Model Bias & Interpretation Gaps

Bias Risk

AI models may underperform on underrepresented patient populations. Always cross-check critical outputs with clinical expertise.

---

🧾 API Error Messages

Here are some typical API errors you may encounter:

Status Code	Meaning	Example Message	Action
400	Bad Request	{"error":"Missing required field: image"}	Check input format and required fields.
401	Unauthorized	{"error":"Invalid API key"}	Verify your API credentials.
429	Too Many Requests	{"error":"Rate limit exceeded"}	Wait or adjust frequency.
500	Internal Server Error	{"error":"Unexpected server failure"}	Retry later. Contact support if persistent.

💡 Where to find the request ID?
Each API response includes a request_id in the headers or JSON body. Include this when contacting support for troubleshooting.

---

🕵️ Debugging Tips

• Log Every Request: Include input, parameters, and timestamp for traceability.
• Compare Runs: Use side-by-side comparisons with slight param changes.
• Leverage debug=true: Enables verbose mode to inspect model decisions and intermediate steps.

---

✅ When to Escalate

Contact Neolens support if:

• Output is empty despite valid input.
• API returns repeated 500 errors.
• Model response is offensive or violates ethical guidelines.

Provide the following when reaching out:

• request_id (from the API response)
• Sample input or image metadata
• Timestamp of the failed request