Update api/server.py

api/server.py

@@ -128,6 +128,11 @@ def _get_session(user_id: str) -> Dict[str, Any]:
             "course_outline": DEFAULT_COURSE_TOPICS,
             "rag_chunks": list(MODULE10_CHUNKS_CACHE),
             "model_name": DEFAULT_MODEL,
+
+            # ✅ NEW: keep track of uploaded files and their chunks
+            "uploaded_chunks_by_file": {},     # Dict[str, List[Dict[str, Any]]]
+            "last_uploaded_filename": None,    # Optional[str]
+            "uploaded_filenames": [],          # List[str]
         }
     return SESSIONS[user_id]
 
@@ -351,32 +356,72 @@ class SummaryReq(BaseModel):
 
 
 class FeedbackReq(BaseModel):
-    # IMPORTANT: allow extra fields so FE can evolve without breaking backend
     class Config:
         extra = "ignore"
 
     user_id: str
     rating: str  # "helpful" | "not_helpful"
-
-    # NEW: attach feedback to a specific LangSmith run
     run_id: Optional[str] = None
-
     assistant_message_id: Optional[str] = None
-
     assistant_text: str
     user_text: Optional[str] = ""
-
     comment: Optional[str] = ""
-
-    # optional structured fields
     tags: Optional[List[str]] = []
     refs: Optional[List[str]] = []
-
     learning_mode: Optional[str] = None
     doc_type: Optional[str] = None
     timestamp_ms: Optional[int] = None
 
 
+# ----------------------------
+# Helpers: prefer last uploaded file when user asks "read/summarize uploaded file"
+# ----------------------------
+def _wants_last_uploaded_file(msg: str) -> bool:
+    t = (msg or "").lower()
+    triggers = [
+        "summarize the uploaded file",
+        "summarise the uploaded file",
+        "summarize uploaded file",
+        "uploaded file",
+        "read this",
+        "can you see that file",
+        "can you see the file",
+        "read the file",
+        "summarize the file i uploaded",
+        "summarize the document i uploaded",
+        "summarize the document",
+        "总结我上传的文件",
+        "总结上传的文件",
+        "读一下我上传的",
+        "能看到我上传的文件吗",
+        "看一下我上传的文件",
+    ]
+    return any(k in t for k in triggers)
+
+
+def _concat_chunks_text(chunks: List[Dict[str, Any]], max_chars: int = 2000) -> str:
+    if not chunks:
+        return ""
+    out: List[str] = []
+    total = 0
+    for c in chunks:
+        # common keys: "text" / "content" / "chunk"
+        txt = c.get("text") or c.get("content") or c.get("chunk") or ""
+        txt = (txt or "").strip()
+        if not txt:
+            continue
+        remain = max_chars - total
+        if remain <= 0:
+            break
+        if len(txt) > remain:
+            txt = txt[:remain]
+        out.append(txt)
+        total += len(txt) + 1
+        if total >= max_chars:
+            break
+    return "\n\n".join(out)
+
+
 # ----------------------------
 # API Routes
 # ----------------------------
@@ -391,6 +436,7 @@ def login(req: LoginReq):
     sess["name"] = name
     return {"ok": True, "user": {"name": name, "user_id": user_id}}
 
+
 @app.post("/api/chat")
 def chat(req: ChatReq):
     user_id = (req.user_id or "").strip()
@@ -411,37 +457,6 @@ def chat(req: ChatReq):
             "run_id": None,
         }
 
-    # ----------------------------
-    # RAG query normalization (short file-intent prompts)
-    # ----------------------------
-    def _looks_like_file_request(text: str) -> bool:
-        t = (text or "").strip().lower()
-        if not t:
-            return False
-        triggers = [
-            "read this",
-            "summarize",
-            "summary",
-            "can you see",
-            "see that file",
-            "see the file",
-            "that file",
-            "this file",
-            "the file",
-            "attached",
-            "attachment",
-            "upload",
-            "uploaded",
-            "document",
-            "pdf",
-            "ppt",
-            "slides",
-            "docx",
-            "analyze",
-            "explain this doc",
-        ]
-        return any(k in t for k in triggers)
-
     t0 = time.time()
     marks_ms: Dict[str, float] = {"start": 0.0}
 
@@ -454,31 +469,27 @@ def chat(req: ChatReq):
     sess["cognitive_state"] = update_cognitive_state_from_message(msg, sess["cognitive_state"])
     marks_ms["cognitive_update_done"] = (time.time() - t0) * 1000.0
 
-    # ✅ Key fix:
-    # - DO NOT gate RAG purely by message length.
-    # - For very short generic messages (e.g. "hi"), skip to save latency.
-    # - For short file-intent messages (e.g. "Read this"), force a better retrieval query.
+    # ✅ RAG selection:
+    # If user explicitly asks to read/summarize the uploaded file, prefer last uploaded file chunks
     rag_context_text, rag_used_chunks = "", []
     try:
-        chunks = sess.get("rag_chunks") or []
-        has_chunks = len(chunks) > 0
-
-        if has_chunks:
-            is_file_intent = _looks_like_file_request(msg)
-            is_too_short_generic = (len(msg) < 8) and (not is_file_intent)
-
-            if is_too_short_generic:
+        if _wants_last_uploaded_file(msg):
+            last_fn = sess.get("last_uploaded_filename")
+            by_file = sess.get("uploaded_chunks_by_file") or {}
+            last_chunks = by_file.get(last_fn) if last_fn else None
+            if last_chunks:
+                rag_context_text = _concat_chunks_text(last_chunks, max_chars=2000)
+                rag_used_chunks = list(last_chunks)[:6]  # keep refs small/stable
+            else:
+                # fallback: if no last upload available, do normal retrieval
+                rag_context_text, rag_used_chunks = retrieve_relevant_chunks(msg, sess["rag_chunks"])
+        else:
+            if len(msg) < 20 and ("?" not in msg):
                 rag_context_text, rag_used_chunks = "", []
             else:
-                retrieval_query = msg
-                if is_file_intent:
-                    # Make retrieval robust even when user message is vague.
-                    # Include doc_type to bias retrieval toward recently uploaded material.
-                    retrieval_query = f"uploaded document ({req.doc_type}) key content and relevant excerpts for: {msg}"
-
-                rag_context_text, rag_used_chunks = retrieve_relevant_chunks(retrieval_query, chunks)
+                rag_context_text, rag_used_chunks = retrieve_relevant_chunks(msg, sess["rag_chunks"])
     except Exception as e:
-        print(f"[chat] rag retrieve error: {repr(e)}")
+        print(f"[chat] rag error: {repr(e)}")
         rag_context_text, rag_used_chunks = "", []
 
     marks_ms["rag_retrieve_done"] = (time.time() - t0) * 1000.0
@@ -572,15 +583,11 @@ def quiz_start(req: QuizStartReq):
 
     sess = _get_session(user_id)
 
-    # 用 quiz instruction 启动（不更新 weaknesses/cognitive_state，避免“系统指令”污染状态）
     quiz_instruction = MICRO_QUIZ_INSTRUCTION
-
     t0 = time.time()
 
-    # 语言：如果 Auto，让 detect_language 决定；否则按传入语言
     resolved_lang = detect_language(quiz_instruction, req.language_preference)
 
-    # RAG：强制用 module10/当前 session 的 chunks，检索一个稳定 query
     rag_context_text, rag_used_chunks = retrieve_relevant_chunks(
         "Module 10 quiz", sess["rag_chunks"]
     )
@@ -588,10 +595,10 @@ def quiz_start(req: QuizStartReq):
     try:
         answer, new_history, run_id = chat_with_clare(
             message=quiz_instruction,
-            history=sess["history"],              # 直接接在当前会话 history 后面
+            history=sess["history"],
             model_name=sess["model_name"],
             language_preference=resolved_lang,
-            learning_mode=req.learning_mode,      # 默认 "quiz"
+            learning_mode=req.learning_mode,
             doc_type=req.doc_type,
             course_outline=sess["course_outline"],
             weaknesses=sess["weaknesses"],
@@ -603,8 +610,6 @@ def quiz_start(req: QuizStartReq):
         return JSONResponse({"error": f"quiz_start failed: {repr(e)}"}, status_code=500)
 
     total_ms = (time.time() - t0) * 1000.0
-
-    # 写回 session history（后续用户回答继续走 /api/chat，会延续 quiz 上下文）
     sess["history"] = new_history
 
     refs = [
@@ -629,7 +634,7 @@ def quiz_start(req: QuizStartReq):
             "refs": refs,
             "rag_used_chunks_count": len(rag_used_chunks or []),
             "history_len": len(sess["history"]),
-            "run_id": run_id,  # NEW
+            "run_id": run_id,
         }
     )
 
@@ -640,7 +645,7 @@ def quiz_start(req: QuizStartReq):
         ),
         "refs": refs,
         "latency_ms": total_ms,
-        "run_id": run_id,  # NEW
+        "run_id": run_id,
     }
 
 
@@ -684,7 +689,19 @@ async def upload(
         print(f"[upload] rag build error: {repr(e)}")
         new_chunks = []
 
-    status_md = f"✅ Loaded base reading + uploaded {doc_type} file."
+    # ✅ NEW: remember this upload as "last uploaded file"
+    try:
+        sess["uploaded_chunks_by_file"] = sess.get("uploaded_chunks_by_file") or {}
+        sess["uploaded_chunks_by_file"][safe_name] = new_chunks
+        sess["last_uploaded_filename"] = safe_name
+        lst = sess.get("uploaded_filenames") or []
+        if safe_name not in lst:
+            lst.append(safe_name)
+        sess["uploaded_filenames"] = lst
+    except Exception as e:
+        print(f"[upload] session remember failed: {repr(e)}")
+
+    status_md = f"✅ Loaded base reading + uploaded {doc_type} file: {safe_name} (chunks={len(new_chunks)})."
 
     _log_event_to_langsmith(
         {
@@ -701,7 +718,7 @@ async def upload(
         }
     )
 
-    return {"ok": True, "added_chunks": len(new_chunks), "status_md": status_md}
+    return {"ok": True, "added_chunks": len(new_chunks), "status_md": status_md, "filename": safe_name}
 
 
 @app.post("/api/feedback")
@@ -717,7 +734,6 @@ def api_feedback(req: FeedbackReq):
     if rating not in ("helpful", "not_helpful"):
         return JSONResponse({"ok": False, "error": "Invalid rating"}, status_code=400)
 
-    # normalize fields
     assistant_text = (req.assistant_text or "").strip()
     user_text = (req.user_text or "").strip()
     comment = (req.comment or "").strip()
@@ -725,7 +741,6 @@ def api_feedback(req: FeedbackReq):
     tags = req.tags or []
     timestamp_ms = int(req.timestamp_ms or int(time.time() * 1000))
 
-    # 1) Dataset event log (what you already have)
     _log_event_to_langsmith(
         {
             "experiment_id": EXPERIMENT_ID,
@@ -736,13 +751,9 @@ def api_feedback(req: FeedbackReq):
             "timestamp_ms": timestamp_ms,
             "rating": rating,
             "assistant_message_id": req.assistant_message_id,
-            "run_id": req.run_id,  # NEW
-
-            # Keep the Example readable:
-            "question": user_text,            # what user asked (optional)
-            "answer": assistant_text,         # the assistant response being rated
-
-            # metadata
+            "run_id": req.run_id,
+            "question": user_text,
+            "answer": assistant_text,
             "comment": comment,
             "tags": tags,
             "refs": refs,
@@ -751,8 +762,6 @@ def api_feedback(req: FeedbackReq):
         }
     )
 
-    # 2) Run-level feedback (attach to actual LangSmith run)
-    #    Only works when FE provides run_id and LangSmith credentials are configured.
     wrote_run_feedback = False
     if req.run_id:
         wrote_run_feedback = _write_feedback_to_langsmith_run(