Added Metrics

app.py

@@ -20,6 +20,7 @@ prometheus_helper = PrometheusHelper()
 
 # --- Build / load index ---
 def build_index():
+    prometheus_helper.start_index_build_timer()
     print("Building FAISS index...")
     ds = load_dataset("olivierdehaene/xkcd", split="train")
     model = SentenceTransformer("all-MiniLM-L6-v2")
@@ -49,6 +50,7 @@ def build_index():
     with open(META_FILE, "wb") as f:
         pickle.dump(meta, f)
 
+    prometheus_helper.stop_index_build_timer()
     return index, meta
 
 def get_index():
@@ -80,10 +82,14 @@ def respond(
 
     prometheus_helper.start_request_timer()
     # Embed the query and search FAISS
+    prometheus_helper.start_faiss_index_search_timer()
     query_vec = embedder.encode([message], convert_to_numpy=True)
     D, I = index.search(query_vec, 5)
     candidates = [meta[int(i)] for i in I[0]]
 
+    prometheus_helper.stop_faiss_index_search_timer()
+    prometheus_helper.start_chat_model_call_timer()
+
     context = "\n".join(
         f"[{c['id']}] {c['title']}\nTranscript: {c['transcript']}\nExplanation: {c['explanation']}"
         for c in candidates
@@ -111,6 +117,8 @@ EXPLANATION
         temperature=0.0, # TODO
     )
 
+    prometheus_helper.stop_chat_model_call_timer()
+
     # Be tolerant to slight schema differences
     try:
         choice = resp.choices[0]
@@ -134,6 +142,7 @@ EXPLANATION
                 img_url = json.load(url)["img"]
                 print(f'Got image url: {img_url}')
             
+            prometheus_helper.record_frequency(int(id))
             return [out_text, gr.Image(value=img_url)]
         except ValueError:
             print("Couldn't parse xkcd ID or get image! That should not happen.")

prometheus_helper.py

@@ -1,4 +1,4 @@
-from prometheus_client import start_http_server, Counter, Summary
+from prometheus_client import start_http_server, Counter, Summary, Histogram
 import time
 
 class PrometheusHelper:
@@ -7,6 +7,27 @@ class PrometheusHelper:
         self.successful_requests_counter = Counter('app_successful_requests_total', 'Total number of successful requests')
         self.failed_requests_counter = Counter('app_failed_requests_total', 'Total number of failed requests')
         self.request_duration_summary = Summary('app_request_duration_seconds', 'Time spent processing request')
+        self.comic_frequency = Counter('comic_frequency', 'Frequency of comics being selected', ['comic_id'])
+        self.index_build_start_time = None
+        self.index_build_duration = Summary('index_build_duration_seconds', 'Time spent building the index')
+        self.faiss_index_search_start_time = None
+        self.faiss_index_search_duration = Summary('faiss_index_search_duration_seconds', 'Time spent searching the index')
+        self.chat_model_call_start_time = None
+        self.chat_model_call_duration = Summary('chat_model_call_duration_seconds', 'Time spent calling the chat model')
+        # self.frequency_histogram = Histogram('comic_frequency', 'Frequency of comics being selected')
+
+
+    def start_faiss_index_search_timer(self):
+        self.faiss_index_search_start_time = time.time()
+    
+    def stop_faiss_index_search_timer(self):
+        self.faiss_index_search_duration.observe(time.time() - self.faiss_index_search_start_time)
+
+    def start_chat_model_call_timer(self):
+        self.chat_model_call_start_time = time.time()
+    
+    def stop_chat_model_call_timer(self):
+        self.chat_model_call_duration.observe(time.time() - self.chat_model_call_start_time)
 
     def record_request(self, success: bool):
         self.request_counter.inc()
@@ -15,6 +36,12 @@ class PrometheusHelper:
         else:
             self.failed_requests_counter.inc()
 
+    def start_index_build_timer(self):
+        self.index_build_start_time = time.time()
+
+    def stop_index_build_timer(self):
+        self.index_build_duration.observe(time.time() - self.index_build_start_time)
+
     def setup_prometheus(self):
         start_http_server(8000)
 
@@ -22,4 +49,8 @@ class PrometheusHelper:
         self.start_time = time.time()
     
     def stop_request_timer(self):
-        self.request_duration_summary.observe(time.time() - self.start_time)
+        self.request_duration_summary.observe(time.time() - self.start_time)
+
+    def record_frequency(self, comic_id: int):
+        self.comic_frequency.labels(comic_id=comic_id).inc()
+