Fixed settings to unstructured to make largest chunks the proper size

django_app/pyproject.toml

@@ -80,4 +80,8 @@ build-backend = "poetry.core.masonry.api"
 DJANGO_SETTINGS_MODULE = "redbox_app.settings"
 testpaths = "tests"
 norecursedirs = "tests/test_ai.py"
-
+env_override_existing_values = 1
+env_files = [
+    ".env.test",
+    ".env"
+]

redbox-core/redbox/chains/ingest.py

@@ -3,12 +3,13 @@
 from io import BytesIO
 from functools import partial
 
+from annotated_types import doc
 from langchain.vectorstores import VectorStore
 from langchain_core.documents.base import Document
 from langchain_core.runnables import RunnableLambda, chain, Runnable
 
+from redbox.loader.loaders import UnstructuredChunkLoader
 from redbox.models.settings import Settings
-from redbox.loader.base import BaseRedboxFileLoader
 
 
 if TYPE_CHECKING:
@@ -26,20 +27,20 @@ def log_chunks(chunks: list[Document]):
     return chunks
 
 
-def document_loader(document_loader_type: type[BaseRedboxFileLoader], s3_client: S3Client, env: Settings) -> Runnable:
+def document_loader(document_loader: UnstructuredChunkLoader, s3_client: S3Client, env: Settings) -> Runnable:
     @chain
     def wrapped(file_name: str):
         file_bytes = s3_client.get_object(Bucket=env.bucket_name, Key=file_name)["Body"].read()
-        return document_loader_type(file_name=file_name, file_bytes=BytesIO(file_bytes), env=env).lazy_load()
+        return document_loader.lazy_load(file_name=file_name, file_bytes=BytesIO(file_bytes))
 
     return wrapped
 
 
 def ingest_from_loader(
-    document_loader_type: type[BaseRedboxFileLoader], s3_client: S3Client, vectorstore: VectorStore, env: Settings
+    loader: UnstructuredChunkLoader, s3_client: S3Client, vectorstore: VectorStore, env: Settings
 ) -> Runnable:
     return (
-        document_loader(document_loader_type=document_loader_type, s3_client=s3_client, env=env)
+        document_loader(document_loader=loader, s3_client=s3_client, env=env)
         | RunnableLambda(list)
         | log_chunks
         | RunnableLambda(partial(vectorstore.add_documents, create_index_if_not_exists=False))  # type: ignore[arg-type]

redbox-core/redbox/loader/__init__.py

@@ -1,4 +0,0 @@
-from .loaders import UnstructuredTitleLoader
-from .loaders import UnstructuredLargeChunkLoader
-
-__all__ = ["UnstructuredTitleLoader", "UnstructuredLargeChunkLoader"]

redbox-core/redbox/loader/ingester.py

@@ -6,8 +6,9 @@
 
 from redbox.chains.components import get_embeddings
 from redbox.chains.ingest import ingest_from_loader
-from redbox.loader import UnstructuredLargeChunkLoader, UnstructuredTitleLoader
+from redbox.loader.loaders import UnstructuredChunkLoader
 from redbox.models import Settings
+from redbox.models.file import ChunkResolution
 
 if TYPE_CHECKING:
     from mypy_boto3_s3.client import S3Client
@@ -43,14 +44,26 @@ def ingest_file(file_name: str) -> str | None:
     es.indices.create(index=es_index_name, ignore=[400])
 
     chunk_ingest_chain = ingest_from_loader(
-        document_loader_type=UnstructuredTitleLoader,
+        loader=UnstructuredChunkLoader(
+            chunk_resolution=ChunkResolution.normal,
+            env=env,
+            min_chunk_size=env.worker_ingest_min_chunk_size,
+            max_chunk_size=env.worker_ingest_max_chunk_size,
+            overlap_chars=0
+        ),
         s3_client=env.s3_client(),
         vectorstore=get_elasticsearch_store(es, es_index_name),
         env=env,
     )
 
     large_chunk_ingest_chain = ingest_from_loader(
-        document_loader_type=UnstructuredLargeChunkLoader,
+        loader=UnstructuredChunkLoader(
+            chunk_resolution=ChunkResolution.largest,
+            env=env,
+            min_chunk_size=env.worker_ingest_largest_chunk_size,
+            max_chunk_size=env.worker_ingest_largest_chunk_size,
+            overlap_chars=env.worker_ingest_largest_chunk_overlap
+        ),
         s3_client=env.s3_client(),
         vectorstore=get_elasticsearch_store_without_embeddings(es, es_index_name),
         env=env,

redbox-core/redbox/loader/loaders.py

@@ -1,13 +1,14 @@
 from collections.abc import Iterator
 from datetime import UTC, datetime
+from io import BytesIO
 from typing import TYPE_CHECKING
 
 import requests
 import tiktoken
 from langchain_core.documents import Document
 
 from redbox.models.file import ChunkResolution, ChunkMetadata
-from redbox.loader.base import BaseRedboxFileLoader
+from redbox.models.settings import Settings
 
 encoding = tiktoken.get_encoding("cl100k_base")
 
@@ -17,76 +18,38 @@
     S3Client = object
 
 
-class UnstructuredLargeChunkLoader(BaseRedboxFileLoader):
+class UnstructuredChunkLoader:
     """Load, partition and chunk a document using local unstructured library"""
 
-    def lazy_load(self) -> Iterator[Document]:  # <-- Does not take any arguments
-        """A lazy loader that reads a file line by line.
-
-        When you're implementing lazy load methods, you should use a generator
-        to yield documents one by one.
-        """
-
-        url = f"http://{self.env.unstructured_host}:8000/general/v0/general"
-        files = {
-            "files": (self.file_name, self.file_bytes),
-        }
-        response = requests.post(
-            url,
-            files=files,
-            data={
-                "strategy": "fast",
-                "max_characters": self.env.worker_ingest_largest_chunk_size,
-                "overlap": self.env.worker_ingest_largest_chunk_overlap,
-                "overlap_all": True,
-            },
-        )
-
-        if response.status_code != 200:
-            raise ValueError(response.text)
-
-        elements = response.json()
+    def __init__(self, chunk_resolution: ChunkResolution, env: Settings, min_chunk_size: int, max_chunk_size: int, overlap_chars: int = 0, overlap_all_chunks: bool = True):
+        self.chunk_resolution = chunk_resolution
+        self.env = env
+        self._min_chunk_size = min_chunk_size
+        self._max_chunk_size = max_chunk_size
+        self._overlap_chars = overlap_chars
+        self._overlap_all_chunks = overlap_all_chunks
 
-        if not elements:
-            raise ValueError("Unstructured failed to extract text for this file")
-
-        for i, raw_chunk in enumerate(elements):
-            yield Document(
-                page_content=raw_chunk["text"],
-                metadata=ChunkMetadata(
-                    index=i,
-                    file_name=raw_chunk["metadata"].get("filename"),
-                    page_number=raw_chunk["metadata"].get("page_number"),
-                    created_datetime=datetime.now(UTC),
-                    token_count=len(encoding.encode(raw_chunk["text"])),
-                    chunk_resolution=ChunkResolution.largest,
-                ).model_dump(),
-            )
-
-
-class UnstructuredTitleLoader(BaseRedboxFileLoader):
-    """Load, partition and chunk a document using local unstructured library"""
-
-    def lazy_load(self) -> Iterator[Document]:  # <-- Does not take any arguments
+    def lazy_load(self, file_name: str, file_bytes: BytesIO) -> Iterator[Document]:
         """A lazy loader that reads a file line by line.
 
         When you're implementing lazy load methods, you should use a generator
         to yield documents one by one.
         """
 
         url = f"http://{self.env.unstructured_host}:8000/general/v0/general"
-
         files = {
-            "files": (self.file_name, self.file_bytes),
+            "files": (file_name, file_bytes),
         }
         response = requests.post(
             url,
             files=files,
             data={
-                "chunking_strategy": "by_title",
                 "strategy": "fast",
-                "combine_under_n_chars": self.env.worker_ingest_min_chunk_size,
-                "max_characters": self.env.worker_ingest_max_chunk_size,
+                "chunking_strategy": "by_title",
+                "max_characters": self._max_chunk_size,
+                "combine_under_n_chars": self._min_chunk_size,
+                "overlap": self._overlap_chars,
+                "overlap_all": self._overlap_all_chunks,
             },
         )
 
@@ -107,6 +70,6 @@ def lazy_load(self) -> Iterator[Document]:  # <-- Does not take any arguments
                     page_number=raw_chunk["metadata"].get("page_number"),
                     created_datetime=datetime.now(UTC),
                     token_count=len(encoding.encode(raw_chunk["text"])),
-                    chunk_resolution=ChunkResolution.normal,
+                    chunk_resolution=self.chunk_resolution,
                 ).model_dump(),
-            )
+            )

redbox-core/tests/test_ingest.py

@@ -14,9 +14,7 @@
 from redbox.loader import ingester
 from redbox.loader.ingester import ingest_file
 from redbox.chains.ingest import document_loader, ingest_from_loader
-from redbox.loader.base import BaseRedboxFileLoader
-from redbox.loader.loaders import UnstructuredTitleLoader
-from redbox.loader.loaders import UnstructuredLargeChunkLoader
+from redbox.loader.loaders import UnstructuredChunkLoader
 from redbox.models.settings import Settings
 from redbox.retriever.queries import make_query_filter
 from redbox.models.file import ChunkResolution
@@ -49,12 +47,8 @@ def make_file_query(file_name: str, resolution: ChunkResolution | None = None) -
 
 
 @patch("redbox.loader.loaders.requests.post")
-@pytest.mark.parametrize(
-    "document_loader_type",
-    [UnstructuredTitleLoader, UnstructuredLargeChunkLoader],
-)
 def test_document_loader(
-    mock_post: MagicMock, document_loader_type: type[BaseRedboxFileLoader], s3_client: S3Client, env: Settings
+    mock_post: MagicMock, s3_client: S3Client, env: Settings
 ):
     """
     Given that I have written a text File to s3
@@ -77,26 +71,31 @@ def test_document_loader(
             },
         }
     ]
+    loader = UnstructuredChunkLoader(
+        chunk_resolution=ChunkResolution.normal,
+        env=env,
+        min_chunk_size=env.worker_ingest_min_chunk_size,
+        max_chunk_size=env.worker_ingest_max_chunk_size
+    )
 
     # Upload file and and call
     file = file_to_s3("html/example.html", s3_client, env)
-    loader = document_loader(document_loader_type, s3_client, env)
+    loader = document_loader(loader, s3_client, env)
     chunks = list(loader.invoke(file))
 
     assert len(chunks) > 0
 
 
 @patch("redbox.loader.loaders.requests.post")
 @pytest.mark.parametrize(
-    "document_loader_type, resolution, has_embeddings",
+    "resolution, has_embeddings",
     [
-        (UnstructuredTitleLoader, ChunkResolution.normal, True),
-        (UnstructuredLargeChunkLoader, ChunkResolution.largest, False),
+        (ChunkResolution.normal, True),
+        (ChunkResolution.largest, False),
     ],
 )
 def test_ingest_from_loader(
     mock_post: MagicMock,
-    document_loader_type: type[BaseRedboxFileLoader],
     resolution: ChunkResolution,
     has_embeddings: bool,
     monkeypatch: MonkeyPatch,
@@ -127,13 +126,20 @@ def test_ingest_from_loader(
         }
     ]
 
+    loader = UnstructuredChunkLoader(
+        chunk_resolution=resolution,
+        env=env,
+        min_chunk_size=env.worker_ingest_min_chunk_size,
+        max_chunk_size=env.worker_ingest_max_chunk_size
+    )
+
     # Mock embeddings
     monkeypatch.setattr(ingester, "get_embeddings", lambda _: FakeEmbeddings(size=3072))
 
     # Upload file and call
     file_name = file_to_s3(filename="html/example.html", s3_client=s3_client, env=env)
     ingest_chain = ingest_from_loader(
-        document_loader_type=document_loader_type, s3_client=s3_client, vectorstore=es_vector_store, env=env
+        loader=loader, s3_client=s3_client, vectorstore=es_vector_store, env=env
     )
 
     _ = ingest_chain.invoke(file_name)