From 1ac6879268e3526bf71006b46b6cc45d281eda87 Mon Sep 17 00:00:00 2001 From: Patrick Wachter Date: Sat, 22 Mar 2025 13:44:50 +0100 Subject: [PATCH] Add Mistral OCR integration and configuration support --- backend/open_webui/config.py | 5 ++ backend/open_webui/main.py | 2 + backend/open_webui/retrieval/loaders/main.py | 59 +++++++++++++++++++ backend/open_webui/routers/retrieval.py | 16 +++++ backend/requirements.txt | 1 + .../admin/Settings/Documents.svelte | 27 +++++++-- 6 files changed, 105 insertions(+), 5 deletions(-) diff --git a/backend/open_webui/config.py b/backend/open_webui/config.py index 0ac92bd23..02f61696e 100644 --- a/backend/open_webui/config.py +++ b/backend/open_webui/config.py @@ -1727,6 +1727,11 @@ DOCUMENT_INTELLIGENCE_KEY = PersistentConfig( os.getenv("DOCUMENT_INTELLIGENCE_KEY", ""), ) +MISTRAL_OCR_API_KEY = PersistentConfig( + "MISTRAL_OCR_API_KEY", + "rag.mistral_ocr_api_key", + os.getenv("MISTRAL_OCR_API_KEY", ""), +) BYPASS_EMBEDDING_AND_RETRIEVAL = PersistentConfig( "BYPASS_EMBEDDING_AND_RETRIEVAL", diff --git a/backend/open_webui/main.py b/backend/open_webui/main.py index bb78d9003..383523174 100644 --- a/backend/open_webui/main.py +++ b/backend/open_webui/main.py @@ -191,6 +191,7 @@ from open_webui.config import ( DOCLING_SERVER_URL, DOCUMENT_INTELLIGENCE_ENDPOINT, DOCUMENT_INTELLIGENCE_KEY, + MISTRAL_OCR_API_KEY, RAG_TOP_K, RAG_TOP_K_RERANKER, RAG_TEXT_SPLITTER, @@ -582,6 +583,7 @@ app.state.config.TIKA_SERVER_URL = TIKA_SERVER_URL app.state.config.DOCLING_SERVER_URL = DOCLING_SERVER_URL app.state.config.DOCUMENT_INTELLIGENCE_ENDPOINT = DOCUMENT_INTELLIGENCE_ENDPOINT app.state.config.DOCUMENT_INTELLIGENCE_KEY = DOCUMENT_INTELLIGENCE_KEY +app.state.config.MISTRAL_OCR_API_KEY = MISTRAL_OCR_API_KEY app.state.config.TEXT_SPLITTER = RAG_TEXT_SPLITTER app.state.config.TIKTOKEN_ENCODING_NAME = TIKTOKEN_ENCODING_NAME diff --git a/backend/open_webui/retrieval/loaders/main.py b/backend/open_webui/retrieval/loaders/main.py index 295d0414a..e75c69682 100644 --- a/backend/open_webui/retrieval/loaders/main.py +++ b/backend/open_webui/retrieval/loaders/main.py @@ -20,6 +20,9 @@ from langchain_community.document_loaders import ( YoutubeLoader, ) from langchain_core.documents import Document + +from mistralai import Mistral + from open_webui.env import SRC_LOG_LEVELS, GLOBAL_LOG_LEVEL logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL) @@ -163,6 +166,53 @@ class DoclingLoader: raise Exception(f"Error calling Docling: {error_msg}") +class MistralLoader: + def __init__(self, api_key: str, file_path: str): + self.api_key = api_key + self.file_path = file_path + self.client = Mistral(api_key=api_key) + + def load(self) -> list[Document]: + log.info("Uploading file to Mistral OCR") + uploaded_pdf = self.client.files.upload( + file={ + "file_name": self.file_path.split("/")[-1], + "content": open(self.file_path, "rb"), + }, + purpose="ocr", + ) + log.info("File uploaded to Mistral OCR, getting signed URL") + signed_url = self.client.files.get_signed_url(file_id=uploaded_pdf.id) + log.info("Signed URL received, processing OCR") + ocr_response = self.client.ocr.process( + model="mistral-ocr-latest", + document={ + "type": "document_url", + "document_url": signed_url.url, + }, + ) + log.info("OCR processing done, deleting uploaded file") + deleted_pdf = self.client.files.delete(file_id=uploaded_pdf.id) + log.info("Uploaded file deleted") + log.debug("OCR response: %s", ocr_response) + if not hasattr(ocr_response, "pages") or not ocr_response.pages: + log.error("No pages found in OCR response") + return [Document(page_content="No text content found", metadata={})] + + return [ + Document( + page_content=page.markdown, + metadata={ + "page": page.index, + "page_label": page.index + 1, + "total_pages": len(ocr_response.pages), + }, + ) + for page in ocr_response.pages + if hasattr(page, "markdown") and hasattr(page, "index") + ] + + class Loader: def __init__(self, engine: str = "", **kwargs): self.engine = engine @@ -222,6 +272,15 @@ class Loader: api_endpoint=self.kwargs.get("DOCUMENT_INTELLIGENCE_ENDPOINT"), api_key=self.kwargs.get("DOCUMENT_INTELLIGENCE_KEY"), ) + elif ( + self.engine == "mistral_ocr" + and self.kwargs.get("MISTRAL_OCR_API_KEY") != "" + and file_ext + in ["pdf"] # Mistral OCR currently only supports PDF and images + ): + loader = MistralLoader( + api_key=self.kwargs.get("MISTRAL_OCR_API_KEY"), file_path=file_path + ) else: if file_ext == "pdf": loader = PyPDFLoader( diff --git a/backend/open_webui/routers/retrieval.py b/backend/open_webui/routers/retrieval.py index 2bd908606..979cd2c70 100644 --- a/backend/open_webui/routers/retrieval.py +++ b/backend/open_webui/routers/retrieval.py @@ -364,6 +364,9 @@ async def get_rag_config(request: Request, user=Depends(get_admin_user)): "endpoint": request.app.state.config.DOCUMENT_INTELLIGENCE_ENDPOINT, "key": request.app.state.config.DOCUMENT_INTELLIGENCE_KEY, }, + "mistral_ocr_config": { + "api_key": request.app.state.config.MISTRAL_OCR_API_KEY, + }, }, "chunk": { "text_splitter": request.app.state.config.TEXT_SPLITTER, @@ -427,11 +430,16 @@ class DocumentIntelligenceConfigForm(BaseModel): key: str +class MistralOCRConfigForm(BaseModel): + api_key: str + + class ContentExtractionConfig(BaseModel): engine: str = "" tika_server_url: Optional[str] = None docling_server_url: Optional[str] = None document_intelligence_config: Optional[DocumentIntelligenceConfigForm] = None + mistral_ocr_config: Optional[MistralOCRConfigForm] = None class ChunkParamUpdateForm(BaseModel): @@ -553,6 +561,10 @@ async def update_rag_config( request.app.state.config.DOCUMENT_INTELLIGENCE_KEY = ( form_data.content_extraction.document_intelligence_config.key ) + if form_data.content_extraction.mistral_ocr_config is not None: + request.app.state.config.MISTRAL_OCR_API_KEY = ( + form_data.content_extraction.mistral_ocr_config.api_key + ) if form_data.chunk is not None: request.app.state.config.TEXT_SPLITTER = form_data.chunk.text_splitter @@ -659,6 +671,9 @@ async def update_rag_config( "endpoint": request.app.state.config.DOCUMENT_INTELLIGENCE_ENDPOINT, "key": request.app.state.config.DOCUMENT_INTELLIGENCE_KEY, }, + "mistral_ocr_config": { + "api_key": request.app.state.config.MISTRAL_OCR_API_KEY, + }, }, "chunk": { "text_splitter": request.app.state.config.TEXT_SPLITTER, @@ -1007,6 +1022,7 @@ def process_file( PDF_EXTRACT_IMAGES=request.app.state.config.PDF_EXTRACT_IMAGES, DOCUMENT_INTELLIGENCE_ENDPOINT=request.app.state.config.DOCUMENT_INTELLIGENCE_ENDPOINT, DOCUMENT_INTELLIGENCE_KEY=request.app.state.config.DOCUMENT_INTELLIGENCE_KEY, + MISTRAL_OCR_API_KEY=request.app.state.config.MISTRAL_OCR_API_KEY, ) docs = loader.load( file.filename, file.meta.get("content_type"), file_path diff --git a/backend/requirements.txt b/backend/requirements.txt index ca2ea5060..078d8d2d3 100644 --- a/backend/requirements.txt +++ b/backend/requirements.txt @@ -77,6 +77,7 @@ psutil sentencepiece soundfile==0.13.1 azure-ai-documentintelligence==1.0.0 +mistralai==1.6.0 pillow==11.1.0 opencv-python-headless==4.11.0.86 diff --git a/src/lib/components/admin/Settings/Documents.svelte b/src/lib/components/admin/Settings/Documents.svelte index 5ea12c9c0..b105ebdb9 100644 --- a/src/lib/components/admin/Settings/Documents.svelte +++ b/src/lib/components/admin/Settings/Documents.svelte @@ -54,6 +54,8 @@ let documentIntelligenceEndpoint = ''; let documentIntelligenceKey = ''; let showDocumentIntelligenceConfig = false; + let mistralApiKey = ''; + let showMistralOcrConfig = false; let textSplitter = ''; let chunkSize = 0; @@ -189,6 +191,10 @@ toast.error($i18n.t('Document Intelligence endpoint and key required.')); return; } + if (contentExtractionEngine === 'mistral_ocr' && mistralApiKey === '') { + toast.error($i18n.t('Mistral OCR API Key required.')); + return; + } if (!BYPASS_EMBEDDING_AND_RETRIEVAL) { await embeddingModelUpdateHandler(); @@ -220,6 +226,9 @@ document_intelligence_config: { key: documentIntelligenceKey, endpoint: documentIntelligenceEndpoint + }, + mistral_ocr_config: { + api_key: mistralApiKey } } }); @@ -284,6 +293,8 @@ documentIntelligenceEndpoint = res.content_extraction.document_intelligence_config.endpoint; documentIntelligenceKey = res.content_extraction.document_intelligence_config.key; showDocumentIntelligenceConfig = contentExtractionEngine === 'document_intelligence'; + mistralApiKey = res.content_extraction.mistral_ocr_config.api_key; + showMistralOcrConfig = contentExtractionEngine === 'mistral_ocr'; fileMaxSize = res?.file.max_size ?? ''; fileMaxCount = res?.file.max_count ?? ''; @@ -335,21 +346,21 @@
-
+
-
+
{$i18n.t('Content Extraction Engine')}
-
@@ -378,12 +389,18 @@ placeholder={$i18n.t('Enter Document Intelligence Endpoint')} bind:value={documentIntelligenceEndpoint} /> -
+ {:else if contentExtractionEngine === 'mistral_ocr'} +
+ +
{/if}