NL: reject inappropriate words and improve schools (#2952)

* The initial bad-word list is located
[here](https://storage.mtls.cloud.google.com/datcom-website-config/nl_bad_words.txt).

* Also, avoid school types from being considered as stop-words. There
were two other ripple effects to this change:
1. It also requires special-handling for fallback logic (to not say
"schools in sunnyvale" => "sunnyvale")
2. It also requires not regressing the demo query [how big are public
schools in sunnyvale]

Note: making the schools change also uncovered
#2953. This whole
stop-word removal business needs streamlining! Post fishfood maybe, and
as part of fixing 2853.
 
Screenshot


![image](https://github.com/datacommonsorg/website/assets/4375037/55a824e5-2ed3-4358-b928-e421cb9dd99f)

nl_server/gcs.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 BUCKET = 'datcom-nl-models'
-TEMP_DIR = '/tmp/'
 
 import os
 from pathlib import Path
@@ -23,21 +22,15 @@
 from google.cloud import storage
 from sentence_transformers import SentenceTransformer
 
+from shared.lib import gcs as gcs_lib
 
-# Downloads the `embeddings_file` from GCS to TEMP_DIR
-# and return its path.
-def download_embeddings(embeddings_file: str) -> str:
-  storage_client = storage.Client()
-  bucket = storage_client.bucket(bucket_name=BUCKET)
-  blob = bucket.get_blob(embeddings_file)
-  # Download
-  local_embeddings_path = local_path(embeddings_file)
-  blob.download_to_filename(local_embeddings_path)
-  return local_embeddings_path
+
+def download_embeddings(embeddings_filename: str) -> str:
+  return gcs_lib.download_file(bucket=BUCKET, filename=embeddings_filename)
 
 
 def local_path(embeddings_file: str) -> str:
-  return os.path.join(TEMP_DIR, embeddings_file)
+  return os.path.join(gcs_lib.TEMP_DIR, embeddings_file)
 
 
 def download_model_from_gcs(gcs_bucket: Any, local_dir: str,
@@ -78,7 +71,7 @@ def download_model_from_gcs(gcs_bucket: Any, local_dir: str,
 def download_model_folder(model_folder: str) -> str:
   sc = storage.Client()
   bucket = sc.bucket(bucket_name=BUCKET)
-  directory = TEMP_DIR
+  directory = gcs_lib.TEMP_DIR
 
   # Only download if needed.
   model_path = os.path.join(directory, model_folder)

server/__init__.py

@@ -35,6 +35,7 @@
 import server.lib.config as libconfig
 from server.lib.disaster_dashboard import get_disaster_dashboard_data
 import server.lib.i18n as i18n
+from server.lib.nl.common import bad_words
 import server.lib.util as libutil
 import server.services.bigtable as bt
 from server.services.discovery import configure_endpoints_from_ingress
@@ -370,6 +371,7 @@ def create_app():
         secret_response = secret_client.access_secret_version(name=secret_name)
         app.config['PALM_API_KEY'] = secret_response.payload.data.decode(
             'UTF-8')
+    app.config['NL_BAD_WORDS'] = bad_words.load_bad_words()
 
   # Get and save the blocklisted svgs.
   blocklist_svg = []

server/integration_tests/nlnext_test.py

@@ -35,7 +35,8 @@ def run_sequence(self,
                    detector='hybrid',
                    check_place_detection=False,
                    expected_detectors=[],
-                   place_detector='ner'):
+                   place_detector='ner',
+                   failure=''):
     if detector == 'heuristic':
       detection_method = 'Heuristic Based'
     elif detector == 'llm':
@@ -76,6 +77,11 @@ def run_sequence(self,
             }
             infile.write(json.dumps(dbg_to_write, indent=2))
       else:
+        if failure:
+          self.assertTrue(failure in resp["failure"]), resp["failure"]
+          self.assertTrue(not resp["config"])
+          return
+
         if not expected_detectors:
           self.assertTrue(dbg.get('detection_type').startswith(detection_method)), \
             'Query {q} failed!'
@@ -135,13 +141,16 @@ def test_demo_cities_feb2023(self):
     self.run_sequence(
         'demo2_cities_feb2023',
         [
+            # This should list public school entities.
             'How big are the public schools in Sunnyvale',
             'What is the prevalence of asthma there',
             'What is the commute pattern there',
             'How does that compare with San Bruno',
             # Proxy for parks in magiceye
             'Which cities in the SF Bay Area have the highest larceny',
             'What countries in Africa had the greatest increase in life expectancy',
+            # This should list stats about the middle school students.
+            'How many middle schools are there in Sunnyvale',
         ])
 
   def test_demo_fallback(self):
@@ -254,3 +263,9 @@ def test_medium_index(self):
     self.run_sequence('medium_index',
                       ['cars per family in california counties'],
                       idx='medium')
+
+  def test_inappropriate_query(self):
+    self.run_sequence('inappropriate_query',
+                      ['how many wise asses live in sunnyvale?'],
+                      idx='medium',
+                      failure='inappropriate words')