Add tests for RAG prompt

Signed-off-by: Andrew Sy Kim <andrewsy@google.com>

applications/rag/example_notebooks/rag-kaggle-ray-sql-latest.ipynb

@@ -48,7 +48,7 @@
    "outputs": [],
    "source": [
     "# Create a directory to package the contents that need to be downloaded in ray worker\n",
-    "! mkdir -p test"
+    "! mkdir -p rag-app"
    ]
   },
   {
@@ -58,8 +58,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "%%writefile test/test.py\n",
-    "# Comment out the above line if you want to see notebook print out, but the line is required for the actual ray job (the test.py is downloaded by the ray workers)\n",
+    "%%writefile rag-app/job.py\n",
+    "# Comment out the above line if you want to see notebook print out, but the line is required for the actual ray job (the job.py is downloaded by the ray workers)\n",
     "\n",
     "import os\n",
     "import uuid\n",
@@ -276,10 +276,10 @@
     "\n",
     "start_time = time.time()\n",
     "job_id = client.submit_job(\n",
-    "    entrypoint=\"python test.py\",\n",
+    "    entrypoint=\"python job.py\",\n",
     "    # Path to the local directory that contains the entrypoint file.\n",
     "    runtime_env={\n",
-    "        \"working_dir\": \"/home/jovyan/test\", # upload the local working directory to ray workers\n",
+    "        \"working_dir\": \"/home/jovyan/rag-app\", # upload the local working directory to ray workers\n",
     "    }\n",
     ")\n",
     "\n",

applications/rag/tests/ray_job.py

@@ -0,0 +1,192 @@
+# This script is copied from the applications/rag/example_notebooks/rag-kaggle-ray-sql-latest.ipynb for testing purposes
+# TODO: remove this script and execute the notebook directly with nbconvert.
+
+import os
+import uuid
+import ray
+from langchain.document_loaders import ArxivLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from sentence_transformers import SentenceTransformer
+from typing import List
+import torch
+from datasets import load_dataset_builder, load_dataset, Dataset
+from huggingface_hub import snapshot_download
+from google.cloud.sql.connector import Connector, IPTypes
+import sqlalchemy
+
+# initialize parameters
+INSTANCE_CONNECTION_NAME = os.environ["CLOUDSQL_INSTANCE_CONNECTION_NAME"]
+print(f"Your instance connection name is: {INSTANCE_CONNECTION_NAME}")
+DB_NAME = "pgvector-database"
+
+db_username_file = open("/etc/secret-volume/username", "r")
+DB_USER = db_username_file.read()
+db_username_file.close()
+
+db_password_file = open("/etc/secret-volume/password", "r")
+DB_PASS = db_password_file.read()
+db_password_file.close()
+
+# initialize Connector object
+connector = Connector()
+
+# function to return the database connection object
+def getconn():
+    conn = connector.connect(
+        INSTANCE_CONNECTION_NAME,
+        "pg8000",
+        user=DB_USER,
+        password=DB_PASS,
+        db=DB_NAME,
+        ip_type=IPTypes.PRIVATE
+    )
+    return conn
+
+# create connection pool with 'creator' argument to our connection object function
+pool = sqlalchemy.create_engine(
+    "postgresql+pg8000://",
+    creator=getconn,
+)
+
+SHARED_DATA_BASEPATH='/data/rag/st'
+SENTENCE_TRANSFORMER_MODEL = 'intfloat/multilingual-e5-small' # Transformer to use for converting text chunks to vector embeddings
+SENTENCE_TRANSFORMER_MODEL_PATH_NAME='models--intfloat--multilingual-e5-small' # the downloaded model path takes this form for a given model name
+SENTENCE_TRANSFORMER_MODEL_SNAPSHOT="ffdcc22a9a5c973ef0470385cef91e1ecb461d9f" # specific snapshot of the model to use
+SENTENCE_TRANSFORMER_MODEL_PATH = SHARED_DATA_BASEPATH + '/' + SENTENCE_TRANSFORMER_MODEL_PATH_NAME + '/snapshots/' + SENTENCE_TRANSFORMER_MODEL_SNAPSHOT # the path where the model is downloaded one time
+
+# the dataset has been pre-dowloaded to the GCS bucket as part of the notebook in the cell above. Ray workers will find the dataset readily mounted.
+SHARED_DATASET_BASE_PATH="/data/netflix-shows/"
+REVIEWS_FILE_NAME="netflix_titles.csv"
+
+BATCH_SIZE = 100
+CHUNK_SIZE = 1000 # text chunk sizes which will be converted to vector embeddings
+CHUNK_OVERLAP = 10
+TABLE_NAME = 'netflix_reviews_db'  # CloudSQL table name
+DIMENSION = 384  # Embeddings size
+ACTOR_POOL_SIZE = 1 # number of actors for the distributed map_batches function
+
+class Embed:
+  def __init__(self):
+        print("torch cuda version", torch.version.cuda)
+        device="cpu"
+        if torch.cuda.is_available():
+            print("device cuda found")
+            device="cuda"
+
+        print ("reading sentence transformer model from cache path:", SENTENCE_TRANSFORMER_MODEL_PATH)
+        self.transformer = SentenceTransformer(SENTENCE_TRANSFORMER_MODEL_PATH, device=device)
+        self.splitter = RecursiveCharacterTextSplitter(chunk_size=CHUNK_SIZE, chunk_overlap=CHUNK_OVERLAP, length_function=len)
+
+  def __call__(self, text_batch: List[str]):
+      text = text_batch["item"]
+      # print("type(text)=", type(text), "type(text_batch)=", type(text_batch))
+      chunks = []
+      for data in text:
+        splits = self.splitter.split_text(data)
+        # print("len(data)", len(data), "len(splits)=", len(splits))
+        chunks.extend(splits)
+
+      embeddings = self.transformer.encode(
+          chunks,
+          batch_size=BATCH_SIZE
+      ).tolist()
+      print("len(chunks)=", len(chunks), ", len(emb)=", len(embeddings))
+      return {'results':list(zip(chunks, embeddings))}
+
+
+# prepare the persistent shared directory to store artifacts needed for the ray workers
+os.makedirs(SHARED_DATA_BASEPATH, exist_ok=True)
+
+# One time download of the sentence transformer model to a shared persistent storage available to the ray workers
+snapshot_download(repo_id=SENTENCE_TRANSFORMER_MODEL, revision=SENTENCE_TRANSFORMER_MODEL_SNAPSHOT, cache_dir=SHARED_DATA_BASEPATH)
+
+# Process the dataset first, wrap the csv file contents into a Ray dataset
+ray_ds = ray.data.read_csv(SHARED_DATASET_BASE_PATH + REVIEWS_FILE_NAME)
+print(ray_ds.schema)
+
+# Distributed flat map to extract the raw text fields.
+ds_batch = ray_ds.flat_map(lambda row: [{
+    'item': "This is a " + str(row["type"]) + " in " + str(row["country"]) + " called " + str(row["title"]) + 
+    " added at " + str(row["date_added"]) + " whose director is " + str(row["director"]) + 
+    " and with cast: " + str(row["cast"]) + " released at " + str(row["release_year"]) + 
+    ". Its rating is: " + str(row['rating']) + ". Its duration is " + str(row["duration"]) + 
+    ". Its description is " + str(row['description']) + "."
+}])
+print(ds_batch.schema)
+
+# Distributed map batches to create chunks out of each row, and fetch the vector embeddings by running inference on the sentence transformer
+ds_embed = ds_batch.map_batches(
+    Embed,
+    compute=ray.data.ActorPoolStrategy(size=ACTOR_POOL_SIZE),
+    batch_size=BATCH_SIZE,  # Large batch size to maximize GPU utilization.
+    num_gpus=1,  # 1 GPU for each actor.
+    # num_cpus=1,
+)
+
+# Use this block for debug purpose to inspect the embeddings and raw text
+# print("Embeddings ray dataset", ds_embed.schema)
+# for output in ds_embed.iter_rows():
+#     # restrict the text string to be less than 65535
+#     data_text = output["results"][0][:65535]
+#     # vector data pass in needs to be a string  
+#     data_emb = ",".join(map(str, output["results"][1]))
+#     data_emb = "[" + data_emb + "]"
+#     print ("raw text:", data_text, ", emdeddings:", data_emb)
+
+# print("Embeddings ray dataset", ds_embed.schema)
+
+data_text = ""
+data_emb = ""
+
+with pool.connect() as db_conn:
+  db_conn.execute(
+    sqlalchemy.text(
+    "CREATE EXTENSION IF NOT EXISTS vector;"
+    )
+  )
+  db_conn.commit()
+
+  create_table_query = "CREATE TABLE IF NOT EXISTS " + TABLE_NAME + " ( id VARCHAR(255) NOT NULL, text TEXT NOT NULL, text_embedding vector(384) NOT NULL, PRIMARY KEY (id));"
+  db_conn.execute(
+    sqlalchemy.text(create_table_query)
+  )
+  # commit transaction (SQLAlchemy v2.X.X is commit as you go)
+  db_conn.commit()
+  print("Created table=", TABLE_NAME)
+
+  query_text = "INSERT INTO " + TABLE_NAME + " (id, text, text_embedding) VALUES (:id, :text, :text_embedding)"
+  insert_stmt = sqlalchemy.text(query_text)
+  for output in ds_embed.iter_rows():
+    # print ("type of embeddings", type(output["results"][1]), "len embeddings", len(output["results"][1]))
+    # restrict the text string to be less than 65535
+    data_text = output["results"][0][:65535]
+    # vector data pass in needs to be a string  
+    data_emb = ",".join(map(str, output["results"][1]))
+    data_emb = "[" + data_emb + "]"
+    # print("text_embedding is ", data_emb)
+    id = uuid.uuid4()
+    db_conn.execute(insert_stmt, parameters={"id": id, "text": data_text, "text_embedding": data_emb})
+
+  # batch commit transactions
+  db_conn.commit()
+
+  # query and fetch table
+  query_text = "SELECT * FROM " + TABLE_NAME
+  results = db_conn.execute(sqlalchemy.text(query_text)).fetchall()
+  # for row in results:
+  #   print(row)
+
+  # verify results
+  transformer = SentenceTransformer(SENTENCE_TRANSFORMER_MODEL)
+  query_text = "During my holiday in Marmaris we ate here to fit the food. It's really good" 
+  query_emb = transformer.encode(query_text).tolist()
+  query_request = "SELECT id, text, text_embedding, 1 - ('[" + ",".join(map(str, query_emb)) + "]' <=> text_embedding) AS cosine_similarity FROM " + TABLE_NAME + " ORDER BY cosine_similarity DESC LIMIT 5;" 
+  query_results = db_conn.execute(sqlalchemy.text(query_request)).fetchall()
+  db_conn.commit()
+  print("print query_results, the 1st one is the hit")
+  for row in query_results:
+    print(row)
+
+# cleanup connector object
+connector.close()
+print ("end job")

applications/rag/tests/test_frontend.py

@@ -6,6 +6,9 @@ def test_frontend_up(rag_frontend_url):
     r.raise_for_status()
     print("Rag frontend is up.")
 
+    assert "Submit your query below." in r.content.decode('utf-8')
+    assert "Enable Filters" in r.content.decode('utf-8')
+
 hub_url = "http://" + sys.argv[1]
 
 test_frontend_up(hub_url)

applications/rag/tests/test_rag.py

@@ -0,0 +1,65 @@
+import json
+import sys
+import requests
+
+def test_prompts(prompt_url):
+    testcases = [
+        {
+            "prompt": "List the cast of Squid Game",
+            "expected_context": "This is a TV Show in  called Squid Game added at September 17, 2021 whose director is  and with cast: Lee Jung-jae, Park Hae-soo, Wi Ha-jun, Oh Young-soo, Jung Ho-yeon, Heo Sung-tae, Kim Joo-ryoung, Tripathi Anupam, You Seong-joo, Lee You-mi released at 2021. Its rating is: TV-MA. Its duration is 1 Season. Its description is Hundreds of cash-strapped players accept a strange invitation to compete in children's games. Inside, a tempting prize awaits — with deadly high stakes..",
+            "expected_substrings": ["Lee Jung-jae", "Park Hae-soo", "Wi Ha-jun", "Oh Young-soo", "Jung Ho-yeon", "Heo Sung-tae", "Kim Joo-ryoung", "Tripathi Anupam", "You Seong-joo", "Lee You-mi"],
+        },
+        {
+            "prompt": "When was Squid Game released?",
+            "expected_context": "This is a TV Show in  called Squid Game added at September 17, 2021 whose director is  and with cast: Lee Jung-jae, Park Hae-soo, Wi Ha-jun, Oh Young-soo, Jung Ho-yeon, Heo Sung-tae, Kim Joo-ryoung, Tripathi Anupam, You Seong-joo, Lee You-mi released at 2021. Its rating is: TV-MA. Its duration is 1 Season. Its description is Hundreds of cash-strapped players accept a strange invitation to compete in children's games. Inside, a tempting prize awaits — with deadly high stakes..",
+            "expected_substrings": ["September 17, 2021"],
+        },
+        {
+            "prompt": "What is the rating of Squid Game?",
+            "expected_context": "This is a TV Show in  called Squid Game added at September 17, 2021 whose director is  and with cast: Lee Jung-jae, Park Hae-soo, Wi Ha-jun, Oh Young-soo, Jung Ho-yeon, Heo Sung-tae, Kim Joo-ryoung, Tripathi Anupam, You Seong-joo, Lee You-mi released at 2021. Its rating is: TV-MA. Its duration is 1 Season. Its description is Hundreds of cash-strapped players accept a strange invitation to compete in children's games. Inside, a tempting prize awaits — with deadly high stakes..",
+            "expected_substrings": ["TV-MA"],
+        },
+        {
+            "prompt": "List the cast of Avatar: The Last Airbender",
+            "expected_context": "This is a TV Show in United States called Avatar: The Last Airbender added at May 15, 2020 whose director is  and with cast: Zach Tyler, Mae Whitman, Jack De Sena, Dee Bradley Baker, Dante Basco, Jessie Flower, Mako Iwamatsu released at 2007. Its rating is: TV-Y7. Its duration is 3 Seasons. Its description is Siblings Katara and Sokka wake young Aang from a long hibernation and learn he's an Avatar, whose air-bending powers can defeat the evil Fire Nation..",
+            "expected_substrings": ["Zach Tyler", "Mae Whitman", "Jack De Sena", "Dee Bradley Baker", "Dante Basco", "Jessie Flower", "Mako Iwamatsu"],
+        },
+        {
+            "prompt": "When was Avatar: The Last Airbender added on Netflix?",
+            "expected_context": "This is a TV Show in United States called Avatar: The Last Airbender added at May 15, 2020 whose director is  and with cast: Zach Tyler, Mae Whitman, Jack De Sena, Dee Bradley Baker, Dante Basco, Jessie Flower, Mako Iwamatsu released at 2007. Its rating is: TV-Y7. Its duration is 3 Seasons. Its description is Siblings Katara and Sokka wake young Aang from a long hibernation and learn he's an Avatar, whose air-bending powers can defeat the evil Fire Nation..",
+            "expected_substrings": ["May 15, 2020"],
+        },
+        {
+            "prompt": "What is the rating of Avatar: The Last Airbender?",
+            "expected_context": "This is a TV Show in United States called Avatar: The Last Airbender added at May 15, 2020 whose director is  and with cast: Zach Tyler, Mae Whitman, Jack De Sena, Dee Bradley Baker, Dante Basco, Jessie Flower, Mako Iwamatsu released at 2007. Its rating is: TV-Y7. Its duration is 3 Seasons. Its description is Siblings Katara and Sokka wake young Aang from a long hibernation and learn he's an Avatar, whose air-bending powers can defeat the evil Fire Nation..",
+            "expected_substrings": ["TV-Y7"],
+        },
+    ]
+
+    for testcase in testcases:
+        prompt = testcase["prompt"]
+        expected_context = testcase["expected_context"]
+        expected_substrings = testcase["expected_substrings"]
+
+        print(f"Testing prompt: {prompt}")
+        data = {"prompt": prompt}
+        json_payload = json.dumps(data)
+
+        headers = {'Content-Type': 'application/json'}
+        response = requests.post(prompt_url, data=json_payload, headers=headers)
+        response.raise_for_status()
+
+        response = response.json()
+        context = response['response']['context']
+        text = response['response']['text']
+        user_prompt = response['response']['user_prompt']
+
+        print(f"Reply: {text}")
+
+        assert user_prompt == prompt, f"unexpected user prompt: {user_prompt} != {prompt}"
+        assert context == expected_context, f"unexpected context: {context} != {expected_context}"
+
+        for substring in expected_substrings:
+            assert substring in text, f"substring {substring} not in response:\n {text}"
+
+test_prompts(sys.argv[1])

cloudbuild.yaml

@@ -61,6 +61,7 @@ steps:
         -var=project_id=$PROJECT_ID \
         -var=network_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-$_AUTOPILOT_CLUSTER  \
         -var=subnetwork_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-$_AUTOPILOT_CLUSTER  \
+        -var=subnetwork_region=$_REGION \
         -var=cluster_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-cluster \
         -var=autopilot_cluster=$_AUTOPILOT_CLUSTER \
         -var=cluster_location=$_REGION \
@@ -146,6 +147,7 @@ steps:
         -var-file=workloads-without-iap.example.tfvars \
         -var=project_id=$PROJECT_ID \
         -var=cluster_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-cluster \
+        -var=cluster_location=$_REGION \
         -var=kubernetes_namespace=ml-$SHORT_SHA-$_BUILD_ID \
         -var=workload_identity_service_account=jupyter-sa-$SHORT_SHA-$_BUILD_ID \
         -var=gcs_bucket=gke-aieco-jupyter-$SHORT_SHA-$_BUILD_ID \
@@ -176,6 +178,7 @@ steps:
         -var-file=workloads-without-iap.example.tfvars \
         -var=project_id=$PROJECT_ID \
         -var=cluster_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-cluster \
+        -var=cluster_location=$_REGION \
         -var=kubernetes_namespace=ml-$SHORT_SHA-$_BUILD_ID \
         -var=workload_identity_service_account=jupyter-sa-$SHORT_SHA-$_BUILD_ID \
         -var=gcs_bucket=gke-aieco-jupyter-$SHORT_SHA-$_BUILD_ID \
@@ -210,6 +213,7 @@ steps:
         -var=frontend_add_auth=false \
         -var=project_id=$PROJECT_ID \
         -var=cluster_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-cluster \
+        -var=cluster_location=$_REGION \
         -var=kubernetes_namespace=rag-$SHORT_SHA-$_BUILD_ID \
         -var=gcs_bucket=gke-aieco-rag-$SHORT_SHA-$_BUILD_ID \
         -var=ray_service_account=ray-sa-$SHORT_SHA-$_BUILD_ID \
@@ -249,8 +253,17 @@ steps:
         cd /workspace/applications/rag/tests
         python3 test_frontend.py "127.0.0.1:8081"
         echo "pass" > /workspace/rag_frontend_result.txt
+
+        # Upload locally stored netflix dataset to GCS bucket mounted as /data
+        gsutil cp ./netflix_titles.csv gs://gke-aieco-rag-$SHORT_SHA-$_BUILD_ID/netflix-shows/netflix_titles.csv
+        ray job submit --working-dir . --address=http://127.0.0.1:8265 -- python ray_job.py
+
+        python3 test_rag.py "http://127.0.0.1:8081/prompt"
+        echo "pass" > /workspace/rag_prompt_result.txt
+
     allowFailure: true
     waitFor: ['cleanup jupyterhub', 'cleanup ray cluster']
+    secretEnv: ['KAGGLE_USERNAME', 'KAGGLE_KEY']
 
   - id: 'cleanup rag'
     name: 'gcr.io/$PROJECT_ID/terraform'
@@ -269,6 +282,7 @@ steps:
         -var=frontend_add_auth=false \
         -var=project_id=$PROJECT_ID \
         -var=cluster_name=ml-$SHORT_SHA-$_PR_NUMBER-$_BUILD_ID-cluster \
+        -var=cluster_location=$_REGION \
         -var=kubernetes_namespace=rag-$SHORT_SHA-$_BUILD_ID \
         -var=gcs_bucket=gke-aieco-rag-$SHORT_SHA-$_BUILD_ID \
         -var=ray_service_account=ray-sa-$SHORT_SHA-$_BUILD_ID \
@@ -350,10 +364,15 @@ steps:
           exit 1
         fi
 
+        if [[ $(cat /workspace/rag_prompt_result.txt) != "pass" ]]; then
+          echo "rag frontend test failed"
+          exit 1
+        fi
+
     waitFor: ['cleanup gke cluster']
 
 substitutions:
-  _REGION: us-central1
+  _REGION: us-east4
   _USER_NAME: github
   _AUTOPILOT_CLUSTER: "false"
   _BUILD_ID: ${BUILD_ID:0:8}

infrastructure/tfvars_tests/standard-gke-public.platform.tfvars

@@ -50,10 +50,9 @@ cpu_pools = [{
 ## make sure required gpu quotas are available in the corresponding region
 enable_gpu = true
 gpu_pools = [{
-  name           = "gpu-pool-l4"
-  machine_type   = "g2-standard-24"
-  node_locations = "us-central1-a"
-  autoscaling    = true
+  name         = "gpu-pool-l4"
+  machine_type = "g2-standard-24"
+  autoscaling  = true
 
   min_count          = 2
   max_count          = 3
@@ -64,4 +63,4 @@ gpu_pools = [{
   disk_type          = "pd-balanced"
   accelerator_type   = "nvidia-l4"
   gpu_driver_version = "LATEST"
-}]
+}]