Понимание документов с помощью Mistral OCR и Milvus

В этом руководстве показано, как построить систему понимания документов с помощью:

Mistral OCR

Мощный сервис оптического распознавания символов, который:

• обрабатывает PDF-файлы, изображения и другие форматы документов
• Сохраняет структуру и форматирование документа
• Работает с многостраничными документами
• Распознает таблицы, списки и другие сложные элементы.

Mistral Embeddings

• Преобразование текста в числовые представления:
• Преобразует текст в 1024-мерные векторы
• Улавливает семантические связи между понятиями
• Обеспечивает сопоставление по сходству на основе смысла
• Обеспечивает основу для семантического поиска

Векторная база данных Milvus

Специализированная база данных для поиска векторного сходства:

• Open-Source
• Выполняет эффективный векторный поиск
• Масштабируется до больших коллекций документов
• Поддерживает гибридный поиск (векторное сходство + фильтрация метаданных)
• Оптимизирована для приложений искусственного интеллекта

К концу этого урока у вас будет система, которая сможет

1. обрабатывать документы (PDF-файлы/изображения) по URL-адресам
2. Извлекать текст с помощью OCR
3. Хранить текст и векторные вложения в Milvus
4. Выполнять семантический поиск по коллекции документов.

---

Установка и зависимости

Сначала установим необходимые пакеты:

$ pip install mistralai pymilvus python-dotenv

Настройка среды

Вам понадобятся:

1. API-ключ Mistral (получите его на сайте https://console.mistral.ai/).
2. Milvus, запущенный локально через Docker или с помощью Zilliz Cloud.

Давайте настроим наше окружение:

import json
import os
import re

from dotenv import load_dotenv
from mistralai import Mistral
from pymilvus import CollectionSchema, DataType, FieldSchema, MilvusClient
from pymilvus.client.types import LoadState

# Load environment variables from .env file
load_dotenv()

# Initialize clients
api_key = os.getenv("MISTRAL_API_KEY")
if not api_key:
    api_key = input("Enter your Mistral API key: ")
    os.environ["MISTRAL_API_KEY"] = api_key

client = Mistral(api_key=api_key)

# Define models
text_model = "mistral-small-latest"  # For chat interactions
ocr_model = "mistral-ocr-latest"  # For OCR processing
embedding_model = "mistral-embed"  # For generating embeddings

# Connect to Milvus (default: localhost)
milvus_uri = os.getenv("MILVUS_URI", "http://localhost:19530")
milvus_client = MilvusClient(uri=milvus_uri)

# Milvus collection name
COLLECTION_NAME = "document_ocr"

print(f"Connected to Mistral API and Milvus at {milvus_uri}")

Connected to Mistral API and Milvus at http://localhost:19530

Настройка коллекции Milvus

Теперь давайте создадим коллекцию Milvus для хранения данных о документах. Коллекция будет содержать следующие поля:

• id: Первичный ключ (генерируется автоматически)
• url: URL-адрес источника документа
• page_num: Номер страницы в документе
• content: Содержание извлеченного текста
• embedding: Векторное представление содержимого (1024 размера)

def setup_milvus_collection():
    """Create Milvus collection if it doesn't exist."""
    # Check if collection already exists
    if milvus_client.has_collection(COLLECTION_NAME):
        print(f"Collection '{COLLECTION_NAME}' already exists.")
        return

    # Define collection schema
    fields = [
        FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
        FieldSchema(name="url", dtype=DataType.VARCHAR, max_length=500),
        FieldSchema(name="page_num", dtype=DataType.INT64),
        FieldSchema(name="content", dtype=DataType.VARCHAR, max_length=65535),
        FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=1024),
    ]

    schema = CollectionSchema(fields=fields)

    # Create collection
    milvus_client.create_collection(
        collection_name=COLLECTION_NAME,
        schema=schema,
    )

    # Create index for vector search
    index_params = milvus_client.prepare_index_params()
    index_params.add_index(
        field_name="embedding",
        index_type="IVF_FLAT",  # Index type for approximate nearest neighbor search
        metric_type="COSINE",  # Similarity metric
        params={"nlist": 128},  # Number of clusters
    )

    milvus_client.create_index(
        collection_name=COLLECTION_NAME, index_params=index_params
    )

    print(f"Collection '{COLLECTION_NAME}' created successfully with index.")


setup_milvus_collection()

Collection 'document_ocr' already exists.

Основные функции

Давайте реализуем основные функции для нашей системы понимания документов:

# Generate embeddings using Mistral
def generate_embedding(text):
    """Generate embedding for text using Mistral embedding model."""
    response = client.embeddings.create(model=embedding_model, inputs=[text])
    return response.data[0].embedding


# Store OCR results in Milvus
def store_ocr_in_milvus(url, ocr_result):
    """Process OCR results and store in Milvus."""
    # Extract pages from OCR result
    pages = []
    current_page = ""
    page_num = 0

    for line in ocr_result.split("\n"):
        if line.startswith("### Page "):
            if current_page:
                pages.append((page_num, current_page.strip()))
            page_num = int(line.replace("### Page ", ""))
            current_page = ""
        else:
            current_page += line + "\n"

    # Add the last page
    if current_page:
        pages.append((page_num, current_page.strip()))

    # Prepare data for Milvus
    entities = []
    for page_num, content in pages:
        # Generate embedding for the page content
        embedding = generate_embedding(content)

        # Create entity
        entity = {
            "url": url,
            "page_num": page_num,
            "content": content,
            "embedding": embedding,
        }
        entities.append(entity)

    # Insert into Milvus
    if entities:
        milvus_client.insert(collection_name=COLLECTION_NAME, data=entities)
        print(f"Stored {len(entities)} pages from {url} in Milvus.")

    return len(entities)


# Define OCR function
def perform_ocr(url):
    """Apply OCR to a URL (PDF or image)."""
    try:
        # Try PDF OCR first
        response = client.ocr.process(
            model=ocr_model, document={"type": "document_url", "document_url": url}
        )
    except Exception:
        try:
            # If PDF OCR fails, try Image OCR
            response = client.ocr.process(
                model=ocr_model, document={"type": "image_url", "image_url": url}
            )
        except Exception as e:
            return str(e)  # Return error message

    # Format the OCR results
    ocr_result = "\n\n".join(
        [
            f"### Page {i + 1}\n{response.pages[i].markdown}"
            for i in range(len(response.pages))
        ]
    )

    # Store in Milvus
    store_ocr_in_milvus(url, ocr_result)

    return ocr_result


# Process URLs
def process_document(url):
    """Process a document URL and return its contents."""
    print(f"Processing document: {url}")
    ocr_result = perform_ocr(url)
    return ocr_result

Функциональность поиска

Теперь давайте реализуем функцию поиска для получения релевантного содержимого документа:

def search_documents(query, limit=5):
    """Search Milvus for similar content to the query."""
    # Check if collection exists
    if not milvus_client.has_collection(COLLECTION_NAME):
        return "No documents have been processed yet."

    # Load collection if not already loaded
    if milvus_client.get_load_state(COLLECTION_NAME) != LoadState.Loaded:
        milvus_client.load_collection(COLLECTION_NAME)

    print(f"Searching for: {query}")
    query_embedding = generate_embedding(query)

    search_params = {"metric_type": "COSINE", "params": {"nprobe": 10}}

    search_results = milvus_client.search(
        collection_name=COLLECTION_NAME,
        data=[query_embedding],
        anns_field="embedding",
        search_params=search_params,
        limit=limit,
        output_fields=["url", "page_num", "content"],
    )

    results = []

    if not search_results or not search_results[0]:
        return "No matching documents found."

    for i, hit in enumerate(search_results[0]):
        url = hit["entity"]["url"]
        page_num = hit["entity"]["page_num"]
        content = hit["entity"]["content"]
        score = hit["distance"]

        results.append(
            {
                "rank": i + 1,
                "score": score,
                "url": url,
                "page": page_num,
                "content": content[:500] + "..." if len(content) > 500 else content,
            }
        )

    return results


# Get statistics about stored documents
def get_document_stats():
    """Get statistics about documents stored in Milvus."""
    if not milvus_client.has_collection(COLLECTION_NAME):
        return "No documents have been processed yet."

    # Get collection stats
    stats = milvus_client.get_collection_stats(COLLECTION_NAME)
    row_count = stats["row_count"]

    # Get unique URLs
    results = milvus_client.query(
        collection_name=COLLECTION_NAME, filter="", output_fields=["url"], limit=10000
    )

    unique_urls = set()
    for result in results:
        unique_urls.add(result["url"])

    return {
        "total_pages": row_count,
        "unique_documents": len(unique_urls),
        "documents": list(unique_urls),
    }

Демо: Обработка документов

Давайте обработаем несколько примеров документов. Вы можете заменить эти URL на свои собственные документы.

# Example PDF URL (Mistral AI paper)
pdf_url = "https://arxiv.org/pdf/2310.06825.pdf"

# Process the document
ocr_result = process_document(pdf_url)

# Display a preview of the OCR result
print("\nOCR Result Preview:")
print("====================")
print(ocr_result[:1000] + "...")

Processing document: https://arxiv.org/pdf/2310.06825.pdf
Stored 9 pages from https://arxiv.org/pdf/2310.06825.pdf in Milvus.

OCR Result Preview:
====================
### Page 1
# Mistral 7B 

Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed

![img-0.jpeg](img-0.jpeg)


#### Abstract

We introduce Mistral 7B, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7B outperforms the best open 13B model (Llama 2) across all evaluated benchmarks, and the best released 34B model (Llama 1) in reasoning, mathematics, and code generation. Our model leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. We also provide a model fine-tuned to follow instructions, Mistral 7B - Instruct, that surpasses Llama 2 13B - chat mod...

Также обработаем изображение:

# Example image URL (replace with your own)
image_url = "https://s3.eu-central-1.amazonaws.com/readcoop.cis.public-assets.prod/hero/old-german-scripts.png"

# Process the image
try:
    ocr_result = process_document(image_url)
    print("\nImage OCR Result:")
    print("=================")
    print(ocr_result)
except Exception as e:
    print(f"Error processing image: {e}")

Processing document: https://s3.eu-central-1.amazonaws.com/readcoop.cis.public-assets.prod/hero/old-german-scripts.png
Stored 1 pages from https://s3.eu-central-1.amazonaws.com/readcoop.cis.public-assets.prod/hero/old-german-scripts.png in Milvus.

Image OCR Result:
=================
### Page 1
![img-0.jpeg](img-0.jpeg)
![img-1.jpeg](img-1.jpeg)
![img-2.jpeg](img-2.jpeg)
![img-3.jpeg](img-3.jpeg)
![img-4.jpeg](img-4.jpeg)
![img-5.jpeg](img-5.jpeg)
![img-6.jpeg](img-6.jpeg)
![img-7.jpeg](img-7.jpeg)
![img-8.jpeg](img-8.jpeg)
![img-9.jpeg](img-9.jpeg)
![img-10.jpeg](img-10.jpeg)
![img-11.jpeg](img-11.jpeg)
![img-12.jpeg](img-12.jpeg)
![img-13.jpeg](img-13.jpeg)
![img-14.jpeg](img-14.jpeg)
![img-15.jpeg](img-15.jpeg)
![img-16.jpeg](img-16.jpeg)
![img-17.jpeg](img-17.jpeg)
![img-18.jpeg](img-18.jpeg)
![img-19.jpeg](img-19.jpeg)
![img-20.jpeg](img-20.jpeg)
![img-21.jpeg](img-21.jpeg)
![img-22.jpeg](img-22.jpeg)
![img-23.jpeg](img-23.jpeg)
![img-24.jpeg](img-24.jpeg)
![img-25.jpeg](img-25.jpeg)
![img-26.jpeg](img-26.jpeg)
![img-27.jpeg](img-27.jpeg)
![img-28.jpeg](img-28.jpeg)
![img-29.jpeg](img-29.jpeg)
![img-30.jpeg](img-30.jpeg)

Демо: Поиск в документах

Теперь, когда мы обработали несколько документов, давайте выполним поиск по ним:

# Get document statistics
stats = get_document_stats()
print(f"Total pages stored: {stats['total_pages']}")
print(f"Unique documents: {stats['unique_documents']}")
print("\nProcessed documents:")
for i, url in enumerate(stats["documents"]):
    print(f"{i + 1}. {url}")

Total pages stored: 58
Unique documents: 3

Processed documents:
1. https://arxiv.org/pdf/2310.06825.pdf
2. https://s3.eu-central-1.amazonaws.com/readcoop.cis.public-assets.prod/hero/old-german-scripts.png
3. https://arxiv.org/pdf/2410.07073

# Search for information
query = "What is Mistral 7B?"
results = search_documents(query, limit=3)

print(f"Search results for: '{query}'\n")

if isinstance(results, str):
    print(results)
else:
    for result in results:
        print(f"Result {result['rank']} (Score: {result['score']:.2f})")
        print(f"Source: {result['url']} (Page {result['page']})")
        print(f"Content: {result['content']}\n")

Searching for: What is Mistral 7B?
Search results for: 'What is Mistral 7B?'

Result 1 (Score: 0.83)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 2)
Content: Mistral 7B is released under the Apache 2.0 license. This release is accompanied by a reference implementation ${ }^{1}$ facilitating easy deployment either locally or on cloud platforms such as AWS, GCP, or Azure using the vLLM [17] inference server and SkyPilot ${ }^{2}$. Integration with Hugging Face ${ }^{3}$ is also streamlined for easier integration. Moreover, Mistral 7B is crafted for ease of fine-tuning across a myriad of tasks. As a demonstration of its adaptability and superior perform...

Result 2 (Score: 0.83)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 2)
Content: Mistral 7B is released under the Apache 2.0 license. This release is accompanied by a reference implementation ${ }^{1}$ facilitating easy deployment either locally or on cloud platforms such as AWS, GCP, or Azure using the vLLM [17] inference server and SkyPilot ${ }^{2}$. Integration with Hugging Face ${ }^{3}$ is also streamlined for easier integration. Moreover, Mistral 7B is crafted for ease of fine-tuning across a myriad of tasks. As a demonstration of its adaptability and superior perform...

Result 3 (Score: 0.82)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 1)
Content: # Mistral 7B 

Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed

![img-0.jpeg](img-0.jpeg)


#### Abstract

We introduce Mistral 7B, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7...

Попробуйте другой поисковый запрос:

# Search for more specific information
query = "What are the capabilities of Mistral's language models?"
results = search_documents(query, limit=3)

print(f"Search results for: '{query}'\n")

if isinstance(results, str):
    print(results)
else:
    for result in results:
        print(f"Result {result['rank']} (Score: {result['score']:.2f})")
        print(f"Source: {result['url']} (Page {result['page']})")
        print(f"Content: {result['content']}\n")

Searching for: What are the capabilities of Mistral's language models?
Search results for: 'What are the capabilities of Mistral's language models?'

Result 1 (Score: 0.85)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 2)
Content: Mistral 7B is released under the Apache 2.0 license. This release is accompanied by a reference implementation ${ }^{1}$ facilitating easy deployment either locally or on cloud platforms such as AWS, GCP, or Azure using the vLLM [17] inference server and SkyPilot ${ }^{2}$. Integration with Hugging Face ${ }^{3}$ is also streamlined for easier integration. Moreover, Mistral 7B is crafted for ease of fine-tuning across a myriad of tasks. As a demonstration of its adaptability and superior perform...

Result 2 (Score: 0.85)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 2)
Content: Mistral 7B is released under the Apache 2.0 license. This release is accompanied by a reference implementation ${ }^{1}$ facilitating easy deployment either locally or on cloud platforms such as AWS, GCP, or Azure using the vLLM [17] inference server and SkyPilot ${ }^{2}$. Integration with Hugging Face ${ }^{3}$ is also streamlined for easier integration. Moreover, Mistral 7B is crafted for ease of fine-tuning across a myriad of tasks. As a demonstration of its adaptability and superior perform...

Result 3 (Score: 0.84)
Source: https://arxiv.org/pdf/2310.06825.pdf (Page 6)
Content: | Model | Answer |
| :--: | :--: |
| Mistral 7B - Instruct with Mistral system prompt | To kill a Linux process, you can use the `kill' command followed by the process ID (PID) of the process you want to terminate. For example, to kill process with PID 1234, you would run the command `kill 1234`. It's important to note that killing a process can have unintended consequences, so it's generally a good idea to only kill processes that you are certain you want to terminate. Additionally, it's genera...

Заключение

В этом руководстве мы создали полноценную систему понимания документов, используя Mistral OCR и Milvus. Эта система может:

1. обрабатывать документы из URL-адресов
2. Извлекать текст, используя возможности Mistral OCR
3. Генерировать векторные вкрапления для содержимого
4. Хранить как текст, так и векторы в Milvus
5. Выполнять семантический поиск по всем обработанным документам.

Такой подход обеспечивает мощные возможности понимания документов, выходящие за рамки простого подбора ключевых слов, позволяя пользователям находить информацию на основе смысла, а не точного совпадения текста.