Add structured logging for query outputs and update service (#54)

This PR adds logging for the query outputs, so we know which documents
were fed into the context, which model was used and which answer was
returned to the user. Structured logging allows us to parse the logs
later and extract the information we are interested in.

Author: pcmoritz

rag/generate.py

@@ -76,9 +76,10 @@ def get_sources_and_context(query, embedding_model, num_chunks):
                 (embedding, num_chunks),
             )
             rows = cur.fetchall()
+            document_ids = [row[0] for row in rows]
             context = [{"text": row[1]} for row in rows]
             sources = [row[2] for row in rows]
-    return sources, context
+    return document_ids, sources, context
 
 
 class QueryAgent:
@@ -107,7 +108,7 @@ def __init__(
 
     def __call__(self, query, num_chunks=5, stream=True):
         # Get sources and context
-        sources, context = get_sources_and_context(
+        document_ids, sources, context = get_sources_and_context(
             query=query, embedding_model=self.embedding_model, num_chunks=num_chunks
         )
 
@@ -126,6 +127,7 @@ def __call__(self, query, num_chunks=5, stream=True):
         result = {
             "question": query,
             "sources": sources,
+            "document_ids": document_ids,
             "answer": answer,
             "llm": self.llm,
         }

rag/serve.py

@@ -1,11 +1,12 @@
 # You can run the whole script locally with
-# serve run rag.serve:deployment
+# serve run rag.serve:deployment --runtime-env-json='{"env_vars": {"RAY_ASSISTANT_LOGS": "/mnt/shared_storage/ray-assistant-logs/info.log", "RAY_ASSISTANT_SECRET": "ray-assistant-prod"}}'
 
 import json
+import logging
 import os
 import pickle
 from pathlib import Path
-from typing import List
+from typing import Any, Dict, List
 
 import openai
 import ray
@@ -17,6 +18,7 @@
 from slack_bolt import App
 from slack_bolt.adapter.socket_mode import SocketModeHandler
 from starlette.responses import StreamingResponse
+import structlog
 
 from rag.config import MAX_CONTEXT_LENGTHS, ROOT_DIR
 from rag.generate import QueryAgent
@@ -37,7 +39,7 @@ def get_secret(secret_name):
     import boto3
 
     client = boto3.client("secretsmanager", region_name="us-west-2")
-    response = client.get_secret_value(SecretId="ray-assistant")
+    response = client.get_secret_value(SecretId=os.environ["RAY_ASSISTANT_SECRET"])
     return json.loads(response["SecretString"])[secret_name]
 
 
@@ -78,6 +80,17 @@ class Answer(BaseModel):
 @serve.ingress(app)
 class RayAssistantDeployment:
     def __init__(self, num_chunks, embedding_model_name, llm, run_slack=False):
+        # Configure logging
+        logging.basicConfig(filename=os.environ["RAY_ASSISTANT_LOGS"], level=logging.INFO, encoding='utf-8')
+        structlog.configure(
+            processors=[
+                structlog.processors.TimeStamper(fmt="iso"),
+                structlog.processors.JSONRenderer(),
+            ],
+            logger_factory=structlog.stdlib.LoggerFactory(),
+        )
+        self.logger = structlog.get_logger()
+
         # Set credentials
         os.environ["ANYSCALE_API_BASE"] = "https://api.endpoints.anyscale.com/v1"
         os.environ["ANYSCALE_API_KEY"] = get_secret("ANYSCALE_API_KEY")
@@ -111,33 +124,48 @@ def __init__(self, num_chunks, embedding_model_name, llm, run_slack=False):
             self.slack_app = SlackApp.remote()
             self.runner = self.slack_app.run.remote()
 
-    @app.post("/query")
-    def query(self, query: Query) -> Answer:
+    def predict(self, query: Query, stream: bool) -> Dict[str, Any]:
         use_oss_agent = self.router.predict([query.query])[0]
         agent = self.oss_agent if use_oss_agent else self.gpt_agent
-        result = agent(query=query.query, num_chunks=self.num_chunks, stream=False)
+        result = agent(query=query.query, num_chunks=self.num_chunks, stream=stream)
+        return result
+
+    @app.post("/query")
+    def query(self, query: Query) -> Answer:
+        result = self.predict(query, stream=False)
         return Answer.parse_obj(result)
 
-    def produce_streaming_answer(self, result):
+    def produce_streaming_answer(self, query, result):
+        answer = []
         for answer_piece in result["answer"]:
+            answer.append(answer_piece)
             yield answer_piece
+
         if result["sources"]:
             yield "\n\n**Sources:**\n"
             for source in result["sources"]:
                 yield "* " + source + "\n"
 
+        self.logger.info(
+            "finished streaming query",
+            query=query,
+            document_ids=result["document_ids"],
+            llm=result["llm"],
+            answer="".join(answer)
+        )
+
     @app.post("/stream")
     def stream(self, query: Query) -> StreamingResponse:
-        use_oss_agent = self.router.predict([query.query])[0]
-        agent = self.oss_agent if use_oss_agent else self.gpt_agent
-        result = agent(query=query.query, num_chunks=self.num_chunks, stream=True)
+        result = self.predict(query, stream=True)
         return StreamingResponse(
-            self.produce_streaming_answer(result), media_type="text/plain")
+            self.produce_streaming_answer(query.query, result),
+            media_type="text/plain"
+        )
 
 
 # Deploy the Ray Serve app
 deployment = RayAssistantDeployment.bind(
-    num_chunks=7,
+    num_chunks=5,
     embedding_model_name="thenlper/gte-large",
     llm="meta-llama/Llama-2-70b-chat-hf",
 )

rag/service.yaml

@@ -3,4 +3,7 @@ cluster_env: ray-assistant
 ray_serve_config:
   import_path: app.serve:deployment
   runtime_env:
-    working_dir: "https://github.com/ray-project/llm-applications/archive/refs/tags/v0.0.8.zip"
+    working_dir: "https://github.com/ray-project/llm-applications/archive/refs/tags/v0.0.9.zip"
+    env_vars:
+      RAY_ASSISTANT_SECRET: "ray-assistant-prod"
+      RAY_ASSISTANT_LOGS: "/mnt/shared_storage/ray-assistant-logs/info.log"

requirements.txt

@@ -10,6 +10,7 @@ ray
 sentence_transformers
 slack_bolt
 streamlit
+structlog
 typer
 tiktoken