3 related questions

[sharpdima]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

Content

How can I add suggested questions to my RAG system so that it only extracts questions from the top chunks that LLM has gone to, and the answers must be in the text.

[dosubot[bot]]

To add suggested questions in your Dify RAG system so that they're only extracted from the top retrieved chunks and answers are present in the text, enable the feature in your app config with "suggested_questions_after_answer": {"enabled": true}. After generating an answer, use the API endpoint GET /messages/{message_id}/suggested to fetch the suggested questions for display. You can control how many chunks are considered by adjusting the TopK parameter in your knowledge base retrieval settings, ensuring only the top chunks are used for both answer generation and question suggestion.

To make sure suggested questions are strictly answerable from the retrieved text, customize the prompt logic in api/core/llm_generator/prompts.py by editing SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT to instruct the LLM to generate questions only if the answer is explicitly present in the provided context. The extraction and validation process is handled in the RAG pipeline, particularly in the QAIndexProcessor class, so ensure its extraction method is configured to pull both questions and answers from your data. You can also fine-tune retrieval with the Score Threshold parameter for more precise chunk selection. Recent updates allow suggested questions to use variables for better contextual awareness as well.

References:
How to enable and use suggested questions
Prompt customization for suggested questions
QAIndexProcessor extraction details
Recent enhancements and variable support

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[onestardao]

What you’re running into is a pretty common failure mode in RAG systems — the model tends to propose “suggested questions” from a wider set of chunks than intended (semantic drift), instead of strictly limiting itself to the top-k retrieved context. The result is suggestions that look valid, but aren’t actually grounded in the retrieved text.

A few practical checks that usually help:

Make sure your top-k setting and Score Threshold are aligned, so you don’t accidentally allow too wide a retrieval scope.

Add an explicit condition in the prompt logic (e.g. only generate suggestions that appear in the provided context).

For stricter control, tag the retrieved chunks before they enter the LLM — this prevents the model from “inventing” outside of what was fetched.

On our side, we’d classify this as No 11 semantic drift in retrieval. Having a small reproducible test case is often enough to confirm whether it’s the retriever leaking too much, or the LLM not following instructions.

One additional approach some teams use is a kind of semantic firewall: instead of changing infrastructure, you add a lightweight semantic check layer in the pipeline to guarantee the suggestions stay inside the retrieved chunks.

Would you find it useful if I sketched a minimal TXTOS test case you could throw at your setup to confirm where the drift happens?