Fixed

app.py

@@ -101,9 +101,17 @@ def retrieve_context(query, top_k=2):
         logger.error(f"Error in retrieve_context: {str(e)}")
         raise
 
-# Load the full CV at startup
-with open("cv_text.txt", "r", encoding="utf-8") as f:
-    full_cv_text = f.read()
+# Load the full CV at startup with explicit UTF-8 handling
+try:
+    with open("cv_text.txt", "r", encoding="utf-8") as f:
+        full_cv_text = f.read()
+    # Ensure full_cv_text is a string
+    if not isinstance(full_cv_text, str):
+        full_cv_text = str(full_cv_text)
+    logger.info("CV text loaded successfully")
+except Exception as e:
+    logger.error(f"Error loading cv_text.txt: {str(e)}")
+    raise
 
 async def stream_response(query, history):
     logger.info(f"Processing query: {query}")
@@ -121,20 +129,32 @@ async def stream_response(query, history):
         f"CV: {full_cv_text}"
     )
 
+    # Ensure system_prompt is a string and debug its state
+    if not isinstance(system_prompt, str):
+        system_prompt = str(system_prompt)
+    logger.info(f"System prompt type: {type(system_prompt)}, length: {len(system_prompt)}")
+
     # Combine system prompt, history, and current query
     messages = [{"role": "system", "content": system_prompt}] + history + [{"role": "user", "content": query}]
 
     # Estimate token counts and truncate history if necessary
-    system_tokens = len(generator.tokenize(system_prompt))
-    query_tokens = len(generator.tokenize(query))
-    history_tokens = [len(generator.tokenize(msg["content"])) for msg in history]
+    try:
+        system_tokens = len(generator.tokenize(system_prompt.encode('utf-8'), add_bos=True, special=True))
+        query_tokens = len(generator.tokenize(query.encode('utf-8'), add_bos=False, special=True))
+        history_tokens = [len(generator.tokenize(msg["content"].encode('utf-8'), add_bos=False, special=True)) for msg in history]
+    except Exception as e:
+        logger.error(f"Tokenization error: {str(e)}")
+        yield f"data: Sorry, I encountered a tokenization error: {str(e)}\n\n"
+        yield "data: [DONE]\n\n"
+        return
+
     total_tokens = system_tokens + query_tokens + sum(history_tokens) + len(history) * 10 + 10  # Rough estimate for formatting
 
     max_allowed_tokens = generator.n_ctx - 512 - 100  # max_tokens=512, safety_margin=100
 
     while total_tokens > max_allowed_tokens and history:
         removed_msg = history.pop(0)
-        removed_tokens = len(generator.tokenize(removed_msg["content"]))
+        removed_tokens = len(generator.tokenize(removed_msg["content"].encode('utf-8'), add_bos=False, special=True))
         total_tokens -= (removed_tokens + 10)
 
     # Reconstruct messages after possible truncation
@@ -142,21 +162,26 @@ async def stream_response(query, history):
 
     # Generate response with lock
     async with model_lock:
-        for chunk in generator.create_chat_completion(
-            messages=messages,
-            max_tokens=512,
-            stream=True,
-            temperature=0.3,
-            top_p=0.7,
-            repeat_penalty=1.2
-        ):
-            token = chunk['choices'][0]['delta'].get('content', '')
-            if token:
-                if not first_token_logged:
-                    logger.info(f"First token time: {time.time() - start_time:.2f}s")
-                    first_token_logged = True
-                yield f"data: {token}\n\n"
-    yield "data: [DONE]\n\n"
+        try:
+            for chunk in generator.create_chat_completion(
+                messages=messages,
+                max_tokens=512,
+                stream=True,
+                temperature=0.3,
+                top_p=0.7,
+                repeat_penalty=1.2
+            ):
+                token = chunk['choices'][0]['delta'].get('content', '')
+                if token:
+                    if not first_token_logged:
+                        logger.info(f"First token time: {time.time() - start_time:.2f}s")
+                        first_token_logged = True
+                    yield f"data: {token}\n\n"
+            yield "data: [DONE]\n\n"
+        except Exception as e:
+            logger.error(f"Generation error: {str(e)}")
+            yield f"data: Sorry, I encountered an error during generation: {str(e)}\n\n"
+            yield "data: [DONE]\n\n"
 
 class QueryRequest(BaseModel):
     query: str

cv_text.txt

@@ -1,6 +1,7 @@
 Tim Luka Horstmann
 
 E-Mail: [email protected]
+Birthdate: 05 December 2000
 
 EDUCATION
 Sep. 2024–Oct. 2025
@@ -87,6 +88,4 @@ Awards:
 ∙ Honored as one of the best Abitur graduates of NRW, Germany’s largest federal state
 
 Interests:
-Artificial Intelligence | Space Travel | Stock Trading | Piano | Rowing | Tennis | Hiking | Travelling
-
-
+Artificial Intelligence | Space Travel | Stock Trading | Piano | Rowing | Tennis | Hiking | Travelling