Bug report: Multi-turn conversation generation injects empty chunkContent, assistant never sees the source text

Bug report: Multi-turn conversation generation injects empty {{chunkContent}}, assistant never sees the source text

Project: https://github.com/ConardLi/easy-dataset Version: 1.7.2 (observed in the packaged desktop app, .next/server/chunks/5246.js and .next/server/chunks/8197.js)

Summary

When generating multi-turn conversations (DatasetConversations), the assistant and follow-up question prompts are built with an empty {{chunkContent}} even when a valid chunk exists, because of an inverted variable-initialization bug in getAssistantReplyPrompt / getNextQuestionPrompt. The model therefore never sees the source text: it either refuses to answer ("材料中并未包含相关记载" / "no such record in the provided material") or hallucinates quotes that do not exist in the source.

Single-turn QA generation is NOT affected (it passes text: chunk.content directly).

Root cause

In app/api/.../multiTurnConversation prompt helpers (compiled chunks 5246.js / 8197.js, module 75246):

async function c(e, {scenario, roleA, roleB, chunkContent: s, ...}, m = null) {
  let y = "";                                    // BUG: should be `let y = s;`
  if (s.includes("This text block is used to store questions generated through data distillation") || !s) {
    y = "没有可用的参考资料，请根据自己的知识直接生成回复";   // fallback string only set in the EMPTY branch
  }
  return await P$(..., { ..., chunkContent: y, ... });   // normal non-empty chunkContent becomes ""
}

y is initialized to "" and only assigned a value inside the "no material" branch. So:

• chunk non-empty (normal case) → y stays "" → {{chunkContent}} is replaced with an empty string → model sees "## 参考资料:" with nothing after it.
• chunk empty/distillation placeholder → y gets the fallback sentence (this branch is actually correct).

The same bug exists in getNextQuestionPrompt (next-round question generation), so both the reply and the follow-up question are affected.

Impact

• Every multi-turn conversation generated via the UI or the batch task API lacks source context.
• Observed behavior in practice: assistant replies "not found in the provided material" for facts that ARE in the chunk (e.g. 李勣 presented 佩刀/玉带 to 张文瓘's colleagues — the exact passage is in the chunk), or hallucinates details (e.g. invented "扶掖乃行"/"失履" for 苏味道; the source says 味道徒步赴逮，席地菜食).
• Any downstream training/eval dataset built from multi-turn conversations is silently degraded.

Suggested fix

One line per function — initialize y with the chunk content:

let y = s;   // getAssistantReplyPrompt
let y = t;   // getNextQuestionPrompt

or equivalently:

let y = s || (s.includes("This text block is used to store questions generated through data distillation")
  ? "没有可用的参考资料，请根据自己的知识直接生成回复" : "");

Note the two compiled chunks both inline module 75246, so the fix must be applied to both copies (the second one uses variable name i for chunkContent in getNextQuestionPrompt).

Reproduction

1. Create any question bound to a normal chunk.
2. Generate a multi-turn conversation (rounds ≥ 2) via POST /api/projects/:projectId/dataset-conversations.
3. Inspect rawMessages → assistant replies claim the material contains no such record, or quote text that is not in the chunk.
4. Fix applied → assistant quotes the source verbatim (e.g. 佩刀、玉带 / 若子才，无施不可，焉用赠).

Environment

• Easy Dataset 1.7.2 desktop app (Windows)
• Model: qwen3.5-plus (OpenAI-compatible)