ATLAS — Aerospace Technical LLM ASsistant

An AI-powered Retrieval-Augmented Generation (RAG) platform developed during an R&D internship at Thales AVS. ATLAS enables avionics engineers to query large corpora of technical documents (PDFs, Word, Excel, HTML, XML) using natural language, and to automate certification-related engineering tasks.

Context

Built during a generative AI R&D internship at Thales AVS (Valence, France), April 2024 – February 2026. Avionics projects generate thousands of heterogeneous documents; certification requirements (DO-178C, DO-254) produce large volumes of requirements, tests, and traceability documents that engineers must cross-reference constantly. ATLAS was designed to reduce the time spent on manual document search and on repetitive writing tasks.

A full presentation of the system is available in Présentation OPEN.pptx.

Two Operating Modes

Informative — conversational assistant

• Query an LLM-RAG agent that dynamically dispatches questions to multiple specialized RAG systems
• Explains the product, locates documents, surfaces links between requirements
• Cites source documents and pages in every answer for traceability

Generative — engineering task automation

• Analyze a High-Level Requirement (HLR): inputs, outputs, interactions, performance constraints
• Check test coverage of an HLR against existing High-Level Tests (HLT)
• Generate a test architecture from an HLR analysis
• Analyze traceability between HLR and Software Requirements (SR)
• Assist with Change Impact Analysis (CIA), FBL/ABL document updates

Ingestion Pipeline

ATLAS handles two document families:

Unstructured / multimodal (PDF, Word, Excel, HTML)

• PDFs and Word files are parsed with Docling → Markdown, preserving tables, figures, and heading hierarchy; figures are annotated with Mistral when relevant
• Chunking is hybrid: first by heading structure, then by token count (≤ BGE-M3's 8192-token window), with post-processing to avoid splitting tables
• Excel sheets are chunked by row groups (≤ 2 000 tokens), always repeating header rows in each chunk
• HTML files are converted to Markdown with html2text, then split hierarchically by heading and by token with LangChain's RecursiveCharacterTextSplitter (2 000 tokens, 200 overlap)

Structured (XML / TXT requirement trees)

• XML is converted to JSON (lighter for LLMs), then chunked if > 8 000 tokens
• Each chunk always includes the full folder/file path as a prefix and as metadata for filtering

Retrieval

Hybrid RAG — combines two retrieval modes, fused with Reciprocal Rank Fusion (RRF):

• Semantic search — ChromaDB vector store + BGE-M3 embeddings (multilingual, 8 192-token context)
• Keyword search — Whoosh full-text index with BM25F scoring (useful for requirement names, acronyms, identifiers)

Visual RAG — for system design documents (SDD) where formulas and diagrams dominate:

• Page images are indexed with ColSmol 256M (ColPali family, runs on 4 GB VRAM)
• Mistral Medium (multimodal) reads the retrieved page images directly

Agentic Workflow

The conversational assistant is orchestrated by a LangGraph agent that:

1. Dispatches the query to the appropriate RAG systems (hybrid, visual, or folder-tree tool)
2. Aggregates retrieved chunks and page images
3. Generates the final answer with Mistral Medium (temperature = 0, frequency penalty = 0.05)
4. Extracts and returns cited sources and page numbers

Gradio Interface

• Chat interface with conversation history
• Document filter to restrict retrieval to specific files
• Analysis mode selector (informative vs. generative workflow)
• Source viewer: cited pages displayed inline
• HTML export of conversations

Installation

Requirements: Python 3.13, a Mistral API key, a BGE-M3 embedding API, and a GPU with ≥ 4 GB VRAM for visual RAG (optional).

pip install -r requirements.txt

Download the following models from HuggingFace and place them in a models/ directory:

• BAAI/bge-m3 → models/bge-m3/
• colSmol-256M → models/colSmol-256M/ (optional, for visual RAG)
• Docling artifacts → models/docling_artifacts/
• Mistral tokenizer (tekken.json) → models/tekken.json

Then set BASE_PATH in my_paths.py to the absolute path of your installation directory.

Running the interface

python interface_gradio.py

The Gradio UI will be available at http://localhost:7860.

Ingesting documents

Place your raw documents (PDF, Word, Excel, HTML, XML/TXT) in data/Raw_database/, then run:

python update_databases_full_pipeline.py

This runs the full pipeline: Word → PDF conversion, Docling parsing, chunking, figure annotation, vectorization into ChromaDB and Whoosh, and visual RAG embedding. Re-running it is incremental — only new or removed documents are processed.

Tech Stack

Category	Tools
Language	Python 3.13
LLMs	Mistral (small / medium / large)
Embeddings	BGE-M3 (local, via HuggingFace)
Vector store	ChromaDB
Keyword search	Whoosh (BM25F)
Retrieval fusion	Reciprocal Rank Fusion (RRF)
Agentic framework	LangChain + LangGraph
Visual retrieval	ColSmol 256M (ColPali family)
PDF parsing	Docling, PyMuPDF
UI	Gradio
Evaluation	RAGAS (Faithfulness, Answer Relevancy, Context Precision)

Project Structure

ATLAS/
├── my_rag.py                      # Core RAG pipeline (retrieve + answer)
├── my_agentic_rag.py              # Agentic RAG with multi-step reasoning
├── agentic_workflow.py            # High-level workflow orchestration
├── hybrid_search.py               # Hybrid dense + sparse retrieval
├── interface_gradio.py            # Web UI
├── ragas_evaluation.py            # Retrieval evaluation
├── manage_glossary.py             # Domain glossary integration
├── visual_rag_on_sdd.py           # Visual RAG on system design documents
├── process_pdf_documents/         # PDF ingestion pipeline
├── process_html_documents/        # HTML ingestion
├── process_excel_documents/       # Excel ingestion
├── workflows/                     # Specialized engineering workflows
│   ├── requirement_analysis.py
│   ├── workflow_HLR_analysis.py
│   ├── workflow_HLT_coverage.py
│   └── ...
└── requirements.txt