Mini-Rag项目构建-11对话接入rag
814 字
4 分钟
Mini-Rag项目构建-11对话接入rag
项目Github地址: https://github.com/gaodaxiu0406/Mini-Rag
⭐ 如果这个项目对你有帮助,欢迎Github点一个 Star——你的一个星,是对独立开发者最直接的鼓励;发现问题或有想法,也欢迎提 Issue。
对话接入检索
检索到的片段不能直接扔给模型,要把它放进System Prompt,并约束模型只根据资料回答。这是RAG的核心机制————没有这句约束,模型会忽略资料用自己的知识“自由发挥”。
在app.py中新增
def _build_system_prompt(context_texts): if context_texts: ctx = "\n---\n".join(context_texts) return ( "你是一个知识库助手。请仅根据以下资料回答用户问题," "如果资料中没有答案,请如实告知用户,而不要编造答案。\n\n{ctx}" f"【参考资料】\n{ctx}" ) return "你是一个知识库助手。知识库当前为空,请告知用户先上传文档。"改造_stream_chat
def _stream_chat(history_messages, query, top_k=None): # —— ——对话接入检索 加形参top_k if top_k is None: top_k = config.RETRIEVAL_TOP_K
# —— ——对话接入检索 检索相关文档片段 retrieved, scores = rag.retrieve(query, top_k=top_k) context_texts = [c["text"] for c in retrieved] system = _build_system_prompt(context_texts)
# 组装消息:system + 历史消息 + 当前问题 messages = [{"role": "system", "content": system}] for m in history_messages: messages.append({"role": m["role"], "content": m["content"]}) messages.append({"role": "user", "content": query})
try: stream = llm.get_client().chat.completions.create( model=config.DEEPSEEK_MODEL, messages=messages, stream=True, temperature=config.DEEPSEEK_TEMPERATURE, max_tokens=config.DEEPSEEK_MAX_TOKENS, )
# 先发检索元信息 yield f"data: {json.dumps({'type': 'meta', 'retrieved': context_texts, 'scores': scores}, ensure_ascii=False)}\n\n"
for chunk in stream: delta = chunk.choices[0].delta if delta.content: text = delta.content payload = json.dumps({"type": "content", "text": text}, ensure_ascii=False) yield f"data: {payload}\n\n"
yield "data: [DONE]\n\n"
except Exception as e: payload = json.dumps({"type": "error", "message": str(e)}, ensure_ascii=False) yield f"data: {payload}\n\n" yield "data: [DONE]\n\n"在app.py中新增文档的REST接口
@app.route('/api/documents', methods=['GET'])def list_documents(): return jsonify(storage.list_documents())
@app.route('/api/documents/upload', methods=['POST'])def upload_document(): if 'file' not in request.files: return jsonify({"error": "No file part"}), 400 f = request.files['file'] if not f.filename: return jsonify({"error": "No filename"}), 400
content = f.read().decode('utf-8', errors='replace') count = rag.index_document(f.filename, content) return jsonify({"filename": f.filename, "chunk_count": count})
@app.route('/api/documents/<filename>', methods=['DELETE'])def delete_document(filename): storage.remove_document(filename) return jsonify({"deleted": filename})上传接口用request.files接收multipart/form-data,读取原始字节后用decode("utf-8",errors="replace")转为文本————errors="replace"保证即使文件里有非法字符也不会崩。
前端不上文档管理交互
1 顶部DOM引用(放在app.js上方)
const docList = $("doc-list");const fileInput = $("file-input");const uploadStatus = $("upload-status");2 文档管理函数(在sendMessage函数之前插入)
async function loadDocuments() { const docs = await api("GET", "/documents"); docList.innerHTML = ""; docs.forEach((d) => { const li = document.createElement("li"); const name = document.createElement("span"); name.textContent = `${d.filename} (${d.chunk_count}块)`; li.appendChild(name); const del = document.createElement("b"); del.className = "doc-item__close"; del.textContent = "×"; del.addEventListener("click", async (e) => { e.stopPropagation(); await api("DELETE", `/documents/${encodeURIComponent(d.filename)}`); await loadDocuments(); }); li.appendChild(del); docList.appendChild(li); })}
btnUpload.addEventListener("click", () => fileInput.click());
fileInput.addEventListener("change", async () => { const file = fileInput.files[0]; if (!file) return; uploadStatus.textContent = "上传中..."; try { const formData = new FormData(); formData.append("file", file); const res = await fetch(API + "/documents/upload", { method: "POST", body: formData }); const data = await res.json(); if (!res.ok) throw new Error(data.error || `HTTP ${res.status}`); uploadStatus.textContent = `上传成功:${data.filename}(${data.chunk_count}块)`; await loadDocuments(); } catch (error) { uploadStatus.textContent = `上传失败:${error.message}`; } finally { fileInput.value = ""; }});注意文件上传用了原生fetch而非api()————因为要发FormData(multipart/form-data),Content-Type必须由浏览器自动设置(含boundary),不能手动指定application/json。
3 SSE解析中处理meta时间(try块中,error分支后补)
else if (data.type === "meta"){ // 检索元信息,不渲染文本}4 初始化时加载文档列表(修改init())
// 初始加载async function init() { // 刷新会话列表 await loadSessions(); // 刷新文档列表 await loadDocuments();}运行效果
运行app.py,在页面上传知识库文件

支持与分享
如果这篇文章对你有帮助,欢迎分享给更多人或打赏支持!
Mini-Rag项目构建-11对话接入rag
https://gaodaxiu-blog.netlify.app/posts/guide/7项目构建-11对话接入rag/











