fix(prompt): require a non-empty answer body in the RAG system prompt

[EnjoyBacon7]

Problem

A streamed RAG answer can collapse to empty content — e.g. the model emits little more than the mandated [Sources: ...] line, which extract_and_strip_sources_block() then strips to "". This yields a blank assistant turn.

On the next request the client (e.g. Twake) replays that blank turn in the chat history, and the upstream LLM (Mistral via litellm) rejects it:

litellm.BadRequestError: OpenAIException - Invalid assistant message:
role='assistant' content='' tool_calls=None  (HTTP 400)

Observed in production: sending test to a partition triggers the 400 on the follow-up turn.

Fix

Add a rule to sys_prompt_tmpl.txt requiring the model to always produce at least one sentence of text — and, when no source is relevant, to briefly say the documents don't cover the question before ending with [Sources: none] — rather than returning an empty body.

Scope / follow-up

This is a prompt-level mitigation and reduces (does not eliminate) blank turns. The robust fix is a server-side guard that drops/coalesces empty assistant messages before forwarding history to the LLM (the inbound check in routers/openai.py only validates the last message). Tracking that separately.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Ensured the assistant always provides at least one sentence in responses, preventing empty response bodies. Added a fallback explanation when insufficient context is available.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 202cea7a-11e7-4383-beb3-fbf88841b8e5

📥 Commits

Reviewing files that changed from the base of the PR and between a24e83c and 52be26f.

📒 Files selected for processing (1)

• prompts/example1/sys_prompt_tmpl.txt

---

📝 Walkthrough

Walkthrough

A single rule is appended to the system prompt template (prompts/example1/sys_prompt_tmpl.txt) requiring the assistant to always output at least one sentence of text, including a fallback instruction for when the retrieved context does not cover the question.

Changes

System Prompt Non-empty Response Rule

Layer / File(s)	Summary
Non-empty response rule prompts/example1/sys_prompt_tmpl.txt	Adds one explicit rule at line 14 mandating that the response body always contains at least one sentence, with a fallback message when the provided context is insufficient.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

Poem

🐇 A rule so small, just one line long,
To keep replies from going wrong,
When context fails, don't stay mute,
A sentence blooms from every root.
Hop along — no empty stew! 🥕

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title directly and concisely summarizes the main change: enforcing non-empty answer bodies in the RAG system prompt.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/llm-always-answer-prompt

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}