🧠 Synthetic RAG Index

Service to import data from various sources (e.g. PDF, images, Microsoft Office, HTML) and index it in AI Search. Increases data relevance and reduces final size by 90%+. Useful for RAG scenarios with LLM. Hosted in Azure with serverless architecture.

Overview

In a real-world scenario, with a public corpus of 15M characters (222 PDF, 7.330 pages), 2.940 facts were generated (8.41 MB indexed). That's a 93% reduction in document amount compared to the chunck method (48.111 chuncks, 300 characters each).

It includes principles taken from research papers:

1. Repetition removal (https://arxiv.org/abs/2112.11446)
2. Corpus cleaning (https://arxiv.org/abs/1910.10683)
3. Synthetic data generation (https://huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1)

Funcional workflow is as follows:

---
title: Workflow
---
graph LR
  raw[("Raw")]
  sanitize["Sanitize"]
  extract["Extract"]
  chunck["Chunck"]
  synthesis["Synthetisis"]
  page["Page"]
  fact["Fact"]
  critic["Critic"]
  index[("Index")]

  raw --> sanitize
  sanitize --> extract
  extract --> chunck
  chunck --> synthesis
  chunck --> synthesis
  synthesis --> page
  page --> fact
  page --> fact
  fact --> critic
  critic --> index
  critic --> index

Loading

Features

Note

This project is a proof of concept. It is not intended to be used in production. This demonstrates how can be combined Azure serverless technologies and LLM to a high quality search engine for RAG scenarios.

• Cost anything when not used thanks to serverless architecture
• Data can be searched with semantic queries using AI Search
• Deduplicate content
• Extract text from PDF, images, Microsoft Office, HTML
• Garbage data detection
• Index files from more than 1000 pages
• Remove redundant and irrelevant content by synthesis data generation

Format support

Document extraction is based on Azure Document Intelligence, specifically on the prebuilt-layout model. It supports popular formats.

Some formats are first converted to PDF with MuPDF to ensure compatibility with Document Intelligence.

Important

Formats not listed there are treated as binary and decoded with UTF-8 encoding.

Format	OCR	Details
.bmp	✅
.cbz	✅	First converted to PDF with MuPDF.
.docx	✅
.epub	✅	First converted to PDF with MuPDF.
.fb2	✅	First converted to PDF with MuPDF.
.heif	✅
.html	✅
.jpg, .jpeg	✅
.mobi	✅	First converted to PDF with MuPDF.
.pdf	✅	Sanitized & compressed with MuPDF.
.png	✅
.pptx	✅
.svg	✅	First converted to PDF with MuPDF.
.tiff	✅
.xlsx	✅
.xps	✅	First converted to PDF with MuPDF.

Demo

As an example, we take the code_des_assurances_2024_1.pdf file.

First, data is extracted from its binary format:

{
  "created_at": "2024-06-08T19:17:51.229972Z",
  "document_content": "Code des assurances\n===\n\ndroit. org Institut Français d'Information Juridique\n\nDernière modification: 2024-01-01 Edition : 2024-01-19 2347 articles avec 5806 liens 57 références externes\n\nCe code ne contient que du droit positif français, les articles et éléments abrogés ne sont pas inclus. Il est recalculé au fur et à mesure des mises à jour. Pensez à actualiser votre copie régulièrement à partir de codes.droit.org.\n\nCes codes ont pour objectif de démontrer l'utilité de l'ouverture des données publiques juridiques tant législatives que jurisprudentielles. Il s'y ajoute une promotion du mouvement Open Science Juridique avec une incitation au dépôt du texte intégral en accès ouvert des articles de doctrine venant du monde professionnel (Grande Bibliothèque du Droit) et universitaire (HAL-CNRS).\n\nTraitements effectués à partir des données issues des APIs Legifrance et Judilibre. droit.org remercie les acteurs du Web qui autorisent des liens vers leur production : Dictionnaire du Droit Privé (réalisé par MM. Serge Braudo et Alexis Baumann), le Conseil constitutionnel, l'Assemblée Nationale, et le Sénat. [...]",
  "file_path": "raw/code_des_assurances_2024_1.pdf",
  "format": "markdown",
  "langs": ["es", "la", "fr", "ja", "en", "it", "pt", "no"],
  "title": "Code des assurances\n==="
}

Second, document is paged, and each page is synthesized to keep track of the context during all steps:

{
  "synthesis": "The \"Code des assurances\" is structured into several legislative parts and chapters, each dealing with various aspects of insurance law and regulations in France. It covers a wide range of insurance-related subjects including the operation of insurance and reinsurance contracts, the requirements for companies, the obligations of insurers and insured, and the legal framework governing insurance practices. The document includes regulations about the constitution and operation of insurance entities, rules for granting administrative approvals, conditions for opening branches and operating under free provision of services, among others.\n\nSpecifically, it addresses the following:\n1. The legislative basis for insurance contracts.\n2. Detailed provisions on maritime, aerial, and space liability insurances.\n3. Obligations for reporting and transparency in insurance practices.\n4. Rules for life insurance and capitalizations applicable in specific French regions and territories.\n5. Provisions for mandatory insurance types, like vehicle insurance, residence insurance, and insurance of construction work.\n6. Specific rules and exceptions for departments like Bas-Rhin, Haut-Rhin, and Moselle and applicability in French overseas territories. [...]"
}

Third, multiple facts (=Q&A pairs) are generated, and those are critiqued to keep only the most relevant ones:

{
  "facts": [
    {
      "answer": "The 'Code des assurances' only contains active French law; abrogated articles and elements are not included.",
      "context": "This exclusion ensures that the code remains up-to-date and relevant, reflecting the current legal landscape without outdated information.",
      "question": "What elements are excluded from the 'Code des assurances'?"
    },
    {
      "answer": "Insurance can be contracted for the policyholder, for another specified person, or for whomever it may concern.",
      "context": "This flexibility allows insurance policies to be tailored to various scenarios, ensuring broad applicability and relevance to different stakeholders.",
      "question": "For whom can insurance be contracted according to the document?"
    }
  ]
}

Finally, facts are individually indexed in AI Search:

{
  "answer": "The 'Code des assurances' only contains active French law; abrogated articles and elements are not included.",
  "context": "This exclusion ensures that the code remains up-to-date and relevant, reflecting the current legal landscape without outdated information.",
  "document_synthesis": "The \"Code des assurances\" is structured into several legislative parts and chapters, each dealing with various aspects of insurance law and regulations in France. It covers a wide range of insurance-related subjects including the operation of insurance and reinsurance contracts, the requirements for companies, the obligations of insurers and insured, and the legal framework governing insurance practices. The document includes regulations about the constitution and operation of insurance entities, rules for granting administrative approvals, conditions for opening branches and operating under free provision of services, among others.\n\nSpecifically, it addresses the following:\n1. The legislative basis for insurance contracts.\n2. Detailed provisions on maritime, aerial, and space liability insurances.\n3. Obligations for reporting and transparency in insurance practices.\n4. Rules for life insurance and capitalizations applicable in specific French regions and territories.\n5. Provisions for mandatory insurance types, like vehicle insurance, residence insurance, and insurance of construction work.\n6. Specific rules and exceptions for departments like Bas-Rhin, Haut-Rhin, and Moselle and applicability in French overseas territories. [...]",
  "file_path": "raw/code_des_assurances_2024_1.pdf",
  "id": "93e5846ba121abf6ea3328a7ff5a96b60ab97ce2016166ac0384f2e61a963d6d",
  "question": "What elements are excluded from the 'Code des assurances'?"
}

High level architecture

---
title: High level process
---
graph LR
  importer["Importer"]
  openai_ada["Ada\n(OpenAI)"]
  search_index["Index\n(AI Search)"]
  storage[("Blob\n(Storage Account)")]

  importer -- Pull from --> storage
  importer -- Push to --> search_index
  search_index -. Generate embeddings .-> openai_ada

Loading

Component level architecture

---
title: Importer component diagram (C4 model)
---
graph LR
  openai_ada["Ada\n(OpenAI)"]
  search_index["Index\n(AI Search)"]
  storage[("Blob\n(Storage Account)")]

  subgraph importer["Importer"]
    document["Document extraction\n(Document Intelligence)"]
    openai_gpt["GPT-4o\n(OpenAI)"]

    func_chunck["Chunck\n(Function App)"]
    func_critic["Critic\n(Function App)"]
    func_extract["Extracted\n(Function App)"]
    func_fact["Fact\n(Function App)"]
    func_index["Index\n(Function App)"]
    func_page["Page\n(Function App)"]
    func_sanitize["Sanitize\n(Function App)"]
    func_synthesis["Synthetisis\n(Function App)"]
  end


  func_sanitize -- Pull from --> storage
  func_sanitize -- Convert and linearize --> func_sanitize
  func_sanitize -- Push to --> func_extract
  func_extract -- Ask for extraction --> document
  func_extract -. Poll for result .-> document
  func_extract -- Push to --> func_chunck
  func_chunck -- Split into large parts --> func_chunck
  func_chunck -- Push to --> func_synthesis
  func_synthesis -- Create a chunck synthesis --> openai_gpt
  func_synthesis -- Push to --> func_page
  func_page -- Split into small parts --> func_page
  func_page -- Clean and filter repetitive content --> func_page
  func_page -- Push to --> func_fact
  func_fact -- Create Q/A pairs --> openai_gpt
  func_fact -- Push to --> func_critic
  func_critic -- Push to --> func_index
  func_critic -- Create a score for each fact --> openai_gpt
  func_critic -- Filter out irrelevant facts --> func_critic
  func_index -- Generate reproductible IDs --> func_index
  func_index -- Push to --> search_index
  search_index -. Generate embeddings .-> openai_ada

Loading

Usage cost

From experiments, the cost of indexing a document is around 29.15€ per 1k pages. Here is a detailed breakdown:

Scenario:

• 7.330 pages (15M characters)
• 222 PDF (550.50 MB)
• French (90%) and English (10%)

Outcome:

• 2.940 facts generated
• 8.41 MB indexed on AI Search

Cost:

Service	Usage	Cost (abs)	Cost (per 1k pages)
Azure AI Search	Billed per hour	N/A	N/A
Azure Blob Storage	N/A	N/A	N/A
Azure Document Intelligence	7.330 pages	67,79€	9.25€
Azure Functions	N/A	N/A	N/A
Azure OpenAI GPT-4o (in)	23.79M tokens	111,81€	15.25€
Azure OpenAI GPT-4o (out)	2.45M tokens	34,06€	4.65€
Total		213,66€	29.15€

Local installation

Prerequisites

Place a file called config.yaml in the root of the project with the following content:

# config.yaml
llm:
  fast:
    mode: azure_openai
    azure_openai:
      api_key: xxx
      context: 16385
      deployment: gpt-35-turbo-0125
      endpoint: https://xxx.openai.azure.com
      model: gpt-35-turbo
      streaming: true
  slow:
    mode: azure_openai
    azure_openai:
      api_key: xxx
      context: 128000
      deployment: gpt-4o-2024-05-13
      endpoint: https://xxx.openai.azure.com
      model: gpt-4o
      streaming: true

destination:
  mode: ai_search
  ai_search:
    access_key: xxx
    endpoint: https://xxx.search.windows.net
    index: trainings

document_intelligence:
  access_key: xxx
  endpoint: https://xxx.cognitiveservices.azure.com

To use a Service Principal to authenticate to Azure, you can also add the following in a .env file:

AZURE_CLIENT_ID=xxx
AZURE_CLIENT_SECRET=xxx
AZURE_TENANT_ID=xxx

To override a specific configuration value, you can also use environment variables. For example, to override the llm.fast.azure_openai.endpoint value, you can use the LLM__FAST__AZURE_OPENAI__ENDPOINT variable:

LLM__FAST__AZURE_OPENAI__ENDPOINT=https://xxx.openai.azure.com

Then run:

# Install dependencies
make install

AI Search also requires to be configured with the following index:

Field Name	Type	Retrievable	Searchable	Dimensions	Vectorizer
answer	Edm.String	Yes	Yes
context	Edm.String	Yes	Yes
created_at	Edm.String	Yes	No
document_synthesis	Edm.String	Yes	Yes
file_path	Edm.String	Yes	No
id	Edm.String	Yes	No
question	Edm.String	Yes	Yes
vectors	Collection(Edm.Single)	No	Yes	1536	OpenAI ADA

Run

Finally, run:

# Start the local API server
make dev

Advanced usage

Configuration

Features are documented in features.py. The features can all be overridden in config.yaml file:

# config.yaml
features:
  fact_iterations: 10
  fact_score_threshold: 0.5
  page_split_size: 2000

[...]