Improve handling of empty retrieval results and API failures in chat responses

[aryarathoree]

Description

When the knowledge-space agent processes a query, there are cases where the retrieval step returns no relevant documents or the LLM / external API call fails or times out. In such situations, the agent currently returns a generic or unclear response, which can be confusing for users.

Expected behavior

• If no relevant documents are retrieved, return a clear, user-friendly message suggesting the user refine or rephrase the query.
• If an LLM or external API call fails, handle the error gracefully and return a meaningful fallback response.
• Log errors internally to aid debugging, without exposing sensitive details to the user.

Proposed improvement

• Add a check for empty retrieval results before invoking the LLM.
• Wrap LLM / API calls in proper error handling.
• Return consistent, descriptive responses for empty results, timeouts, or API failures.

Additional context

This would improve user experience and debuggability, especially for new users unfamiliar with the dataset scope.

[aryarathoree]

I’d be happy to work on this issue if it aligns with the project’s goals.
Thanks!

[tisha-varma]

Hi! I’d like to work on this issue.

My understanding of the problem:

• Empty retrieval results can lead to unclear or unhelpful responses
• API / LLM failures are not handled gracefully in user-facing replies
• Internal logging for these cases is limited or inconsistent

Proposed approach:

1. Add an empty-result check in 'generate_final_response()' (backend/agents.py) before calling the LLM, and return a helpful message suggesting the user refine or rephrase their query.
2. Wrap LLM calls in backend/agents.py with 'try/except' and log errors internally for easier debugging.
3. Improve the generic exception handling in 'chat_endpoint()' (backend/main.py) to return user-friendly fallback messages without exposing internal error details.

Files involved:

• backend/agents.py
• backend/main.py

Quick questions:

• Is there a preferred logging setup (stdlib 'logging' vs existing project conventions)?
• Should the “no results” message reference specific data sources (e.g., DANDI, EBRAINS), or stay generic?

Happy to adjust scope based on feedback. Please let me know if this approach sounds good, and I can proceed with implementation.