Defer AI library imports to reduce process startup cost

learning_resources/converters/opendataloader_llm_converter.py

@@ -33,6 +33,9 @@
 
 log = logging.getLogger(__name__)
 
+# drop unsupported model params
+litellm.drop_params = True
+
 # --- Configuration ---
 MIN_IMAGE_DIMENSION = 32
 MIN_IMAGE_RATIO = 12

learning_resources/etl/utils.py

@@ -46,9 +46,6 @@
     OfferedBy,
     RunStatus,
 )
-from learning_resources.converters.opendataloader_llm_converter import (
-    OpenDataLoaderLLMConverter,
-)
 from learning_resources.etl.constants import (
     RESOURCE_DELIVERY_MAPPING,
     TIME_INTERVAL_MAPPING,
@@ -1294,6 +1291,10 @@ def _pdf_to_markdown(pdf_path):
     """
     Convert a PDF file to markdown using an llm
     """
+    from learning_resources.converters.opendataloader_llm_converter import (
+        OpenDataLoaderLLMConverter,
+    )
+
     converter = OpenDataLoaderLLMConverter(pdf_path)
     return converter.convert_to_markdown()
 

learning_resources/serializers.py

@@ -13,7 +13,6 @@
 from django.db.models import F, Max, Prefetch, Q
 from drf_spectacular.utils import extend_schema_field, extend_schema_serializer
 from isodate import parse_duration
-from langchain_text_splitters import RecursiveJsonSplitter
 from rest_framework import serializers
 from rest_framework.exceptions import ValidationError
 
@@ -1128,6 +1127,8 @@ def render_document(self):
 
     def render_chunks(self):
         """Convert resource info to markdown chunks"""
+        from langchain_text_splitters import RecursiveJsonSplitter
+
         rendered_doc = self.render_document()
         resource_type = self.instance.get("resource_type", "course")
         """

learning_resources/tasks.py

@@ -15,7 +15,6 @@
 from django.utils import timezone
 
 from learning_resources.constants import LearningResourceType
-from learning_resources.content_summarizer import ContentSummarizer
 from learning_resources.etl import ovs, pipelines, youtube
 from learning_resources.etl.canvas import (
     sync_canvas_archive,
@@ -549,6 +548,8 @@ def summarize_content_files_task(
     Returns:
         - None
     """
+    from learning_resources.content_summarizer import ContentSummarizer
+
     summarizer = ContentSummarizer()
     return summarizer.summarize_content_files_by_ids(content_file_ids, overwrite)
 

learning_resources/tasks_test.py

@@ -546,7 +546,7 @@ def test_summarize_unprocessed_content(
         "learning_resources.tasks.summarize_content_files_task", autospec=True
     )
     get_unprocessed_content_file_ids_mock = mocker.patch(
-        "learning_resources.tasks.ContentSummarizer.get_unprocessed_content_file_ids",
+        "learning_resources.content_summarizer.ContentSummarizer.get_unprocessed_content_file_ids",
         autospec=True,
         return_value=ids,
     )

learning_resources/utils.py

@@ -10,7 +10,6 @@
 import html2text
 import rapidjson
 import requests
-import tiktoken
 import yaml
 from botocore.exceptions import ClientError
 from django.conf import settings
@@ -800,6 +799,8 @@ def truncate_to_tokens(text: str, max_tokens: int, model: str = "gpt-4o") -> str
     """
     Truncate text to a maximum number of tokens for a given model.
     """
+    import tiktoken
+
     encoding = tiktoken.encoding_for_model(model)
     tokens = encoding.encode(text)
     if len(tokens) <= max_tokens:

main/sentry.py

@@ -7,6 +7,7 @@
 from sentry_sdk.integrations.celery import CeleryIntegration
 from sentry_sdk.integrations.django import DjangoIntegration
 from sentry_sdk.integrations.logging import LoggingIntegration
+from sentry_sdk.integrations.redis import RedisIntegration
 
 # these errors occur when a shutdown is happening (usually caused by a SIGTERM)
 SHUTDOWN_ERRORS = (WorkerLostError, SystemExit)
@@ -73,9 +74,17 @@ def init_sentry(  # noqa: PLR0913
         before_send=before_send,
         traces_sample_rate=traces_sample_rate,
         profiles_sample_rate=profiles_sample_rate,
+        # Sentry's auto-enabling integrations (langchain, boto3, openai, etc.)
+        # import their target library at init time just to check whether it's
+        # installed, regardless of whether it's ever used -- e.g. langchain is
+        # installed as a litellm transitive dependency but never used directly,
+        # and importing its Sentry integration alone pulls in the whole
+        # langchain/tiktoken stack on every process. Opt in explicitly instead.
+        auto_enabling_integrations=False,
         integrations=[
             DjangoIntegration(),
             CeleryIntegration(),
             LoggingIntegration(level=log_level),
+            RedisIntegration(),
         ],
     )

vector_search/encoders/litellm.py

@@ -13,6 +13,9 @@
 log = logging.getLogger()
 redis_url = urlparse(settings.CELERY_BROKER_URL)
 
+# drop unsupported model params
+litellm.drop_params = True
+
 # these must be set directly via environ (litellm limitation)
 os.environ["REDIS_SSL"] = str(redis_url.scheme.endswith("ss"))
 litellm.cache = Cache(

vector_search/encoders/utils.py

@@ -1,10 +1,8 @@
 import logging
 from functools import cache
 
-import tiktoken
 from django.conf import settings
 from django.utils.module_loading import import_string
-from litellm import get_model_info
 
 logger = logging.getLogger(__name__)
 
@@ -74,6 +72,8 @@ def truncate_to_model_limit(
     Returns:
         Truncated text.
     """
+    from litellm import get_model_info
+
     max_input_tokens = None
 
     try:
@@ -84,6 +84,8 @@ def truncate_to_model_limit(
 
     if max_input_tokens and token_encoding_name:
         try:
+            import tiktoken
+
             encoding = tiktoken.get_encoding(token_encoding_name)
             tokens = encoding.encode(text)
 

vector_search/encoders/utils_test.py

@@ -34,11 +34,11 @@ def warning_mock(mocker):
 def test_truncate_to_model_limit_truncates_long_text(mocker, warning_mock):
     """Text over the model's max input tokens is cut to that token limit"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={"max_input_tokens": 5},
     )
     mocker.patch(
-        "vector_search.encoders.utils.tiktoken.get_encoding",
+        "tiktoken.get_encoding",
         return_value=FakeEncoding(),
     )
 
@@ -55,11 +55,11 @@ def test_truncate_to_model_limit_truncates_long_text(mocker, warning_mock):
 def test_truncate_to_model_limit_leaves_short_text_unchanged(mocker, warning_mock):
     """Text within the model's max input tokens is returned as-is"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={"max_input_tokens": 5},
     )
     mocker.patch(
-        "vector_search.encoders.utils.tiktoken.get_encoding",
+        "tiktoken.get_encoding",
         return_value=FakeEncoding(),
     )
 
@@ -78,12 +78,10 @@ def test_truncate_to_model_limit_unknown_model_falls_back_to_chars(
 ):
     """Fall back to the character limit when model info can't be determined"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         side_effect=ValueError("unknown model"),
     )
-    get_encoding_mock = mocker.patch(
-        "vector_search.encoders.utils.tiktoken.get_encoding"
-    )
+    get_encoding_mock = mocker.patch("tiktoken.get_encoding")
     text = "x" * (DEFAULT_TRUNCATE_CHARS + 100)
 
     result = truncate_to_model_limit(
@@ -102,7 +100,7 @@ def test_truncate_to_model_limit_missing_max_tokens_falls_back_to_chars(
 ):
     """Fall back to the character limit when the model has no max_input_tokens"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={},
     )
     text = "x" * (DEFAULT_TRUNCATE_CHARS + 100)
@@ -122,7 +120,7 @@ def test_truncate_to_model_limit_no_token_encoding_falls_back_to_chars(
 ):
     """Fall back to the character limit when no token encoding name is given"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={"max_input_tokens": 5},
     )
     text = "x" * (DEFAULT_TRUNCATE_CHARS + 100)
@@ -136,11 +134,11 @@ def test_truncate_to_model_limit_no_token_encoding_falls_back_to_chars(
 def test_truncate_to_model_limit_bad_encoding_falls_back_to_chars(mocker, warning_mock):
     """Fall back to the character limit when the tiktoken encoding fails"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={"max_input_tokens": 5},
     )
     mocker.patch(
-        "vector_search.encoders.utils.tiktoken.get_encoding",
+        "tiktoken.get_encoding",
         side_effect=ValueError("bad encoding"),
     )
     text = "x" * (DEFAULT_TRUNCATE_CHARS + 100)
@@ -158,7 +156,7 @@ def test_truncate_to_model_limit_bad_encoding_falls_back_to_chars(mocker, warnin
 def test_truncate_to_model_limit_fallback_warns_once_per_model(mocker, warning_mock):
     """The fallback warning is deduplicated per model/encoding combination"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         side_effect=ValueError("unknown model"),
     )
 
@@ -183,7 +181,7 @@ def test_truncate_to_model_limit_strict_raises_instead_of_falling_back(
 ):
     """In strict mode any fallback condition raises instead of truncating"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         **model_info_kwargs,
     )
     # no token_encoding_name, so even a known model can't use token truncation
@@ -195,11 +193,11 @@ def test_truncate_to_model_limit_strict_raises_instead_of_falling_back(
 def test_truncate_to_model_limit_strict_allows_token_truncation(mocker, warning_mock):
     """Strict mode doesn't interfere when token-based truncation works"""
     mocker.patch(
-        "vector_search.encoders.utils.get_model_info",
+        "litellm.get_model_info",
         return_value={"max_input_tokens": 5},
     )
     mocker.patch(
-        "vector_search.encoders.utils.tiktoken.get_encoding",
+        "tiktoken.get_encoding",
         return_value=FakeEncoding(),
     )
 

vector_search/tasks.py

@@ -10,7 +10,6 @@
 from django.core.cache import caches
 from django.db.models import Q
 
-from learning_resources.content_summarizer import ContentSummarizer
 from learning_resources.models import (
     ContentFile,
     Course,
@@ -741,6 +740,8 @@ def _missing_summaries():
         # of letting it scan every learning resource.
         return []
 
+    from learning_resources.content_summarizer import ContentSummarizer
+
     summarizer = ContentSummarizer()
     return summarizer.get_unprocessed_content_file_ids(
         overwrite=False,

vector_search/utils.py

@@ -8,16 +8,11 @@
 from django.conf import settings
 from django.db import close_old_connections
 from django.db.models import Prefetch, Q
-from langchain_text_splitters import (
-    MarkdownHeaderTextSplitter,
-    RecursiveCharacterTextSplitter,
-)
 from qdrant_client import AsyncQdrantClient, QdrantClient, models
 
 from learning_resources.constants import (
     PROGRAM_COURSE_CACHE_KEY_TEST_MODE,
 )
-from learning_resources.content_summarizer import ContentSummarizer
 from learning_resources.models import (
     ContentFile,
     LearningResource,
@@ -393,6 +388,8 @@ def embed_topics():
 
 @cache
 def _get_text_splitter(**kwargs):
+    from langchain_text_splitters import RecursiveCharacterTextSplitter
+
     if settings.LITELLM_TOKEN_ENCODING_NAME:
         kwargs["encoding_name"] = settings.LITELLM_TOKEN_ENCODING_NAME
         return RecursiveCharacterTextSplitter.from_tiktoken_encoder(**kwargs)
@@ -429,6 +426,8 @@ def _chunk_markdown_documents(text, metadata):
     metadata so every chunk's page_content is self-describing
     for embedding.
     """
+    from langchain_text_splitters import MarkdownHeaderTextSplitter
+
     header_splitter = MarkdownHeaderTextSplitter(
         headers_to_split_on=MARKDOWN_HEADERS_TO_SPLIT_ON,
         strip_headers=False,
@@ -893,6 +892,8 @@ def _summarize_content_files_for_embedding(
     if not fill_summary_content_ids and not changed_summary_content_ids:
         return docs_batch
 
+    from learning_resources.content_summarizer import ContentSummarizer
+
     summarizer = ContentSummarizer()
     if fill_summary_content_ids:
         summarizer.summarize_content_files_by_ids(

vector_search/utils_test.py

@@ -607,7 +607,7 @@ def test_document_chunker_tiktoken(mocker):
     encoder = dense_encoder()
     encoder.token_encoding_name = None
     mocked_splitter = mocker.patch(
-        "vector_search.utils.RecursiveCharacterTextSplitter.from_tiktoken_encoder"
+        "langchain_text_splitters.RecursiveCharacterTextSplitter.from_tiktoken_encoder"
     )
 
     _chunk_documents(["this is a test document"], [{}])
@@ -628,11 +628,15 @@ def test_text_splitter_chunk_size_override(mocker):
     settings.CONTENT_FILE_EMBEDDING_CHUNK_SIZE_OVERRIDE = chunk_size
     settings.CONTENT_FILE_EMBEDDING_CHUNK_OVERLAP = chunk_size / 10
     encoder = dense_encoder()
-    mocked_splitter = mocker.patch("vector_search.utils.RecursiveCharacterTextSplitter")
+    mocked_splitter = mocker.patch(
+        "langchain_text_splitters.RecursiveCharacterTextSplitter"
+    )
     encoder.token_encoding_name = "cl100k_base"  # noqa: S105
     _chunk_documents(["this is a test document"], [{}])
     assert mocked_splitter.mock_calls[0].kwargs["chunk_size"] == 100
-    mocked_splitter = mocker.patch("vector_search.utils.RecursiveCharacterTextSplitter")
+    mocked_splitter = mocker.patch(
+        "langchain_text_splitters.RecursiveCharacterTextSplitter"
+    )
     settings.CONTENT_FILE_EMBEDDING_CHUNK_SIZE_OVERRIDE = None
     _chunk_documents(["this is a test document"], [{}])
     assert "chunk_size" not in mocked_splitter.mock_calls[0].kwargs