Overview

EleAgent is an AI-powered search and query system for civil engineering project documentation. It targets the Karavanke Tunnel renovation project (DARS/Elea iC), replacing manual folder navigation with natural language queries across PDFs, IFC models, and Excel tables.

Phase 1 Goal

Validate the approach with minimal cost. Deploy a Docker image with an AI agent and web UI that gives engineers direct AI-assisted access to project data. No MCP servers, no RAG — conversations are recorded for later analysis to inform Phase 2 tooling.

Why no RAG or MCP in Phase 1?

The AI agent with direct filesystem access and pre-installed libraries is already highly capable for document analysis. Phase 1 prioritizes understanding what engineers actually ask over optimizing response speed. Query times on the order of minutes (vs. seconds with RAG) are acceptable for validation.

What Engineers Can Do

• Ask questions in Slovenian about tunnel specifications, kampada details, material properties, and construction parameters
• Query IFC models — extract PropertySets, element types, geometry data across 276 BIM files
• Analyze Excel tables — cross-reference kampada registries, attribute codebooks, and model inventories
• Search PDFs — extract text and tables from 969 technical documents, including OCR for scanned pages
• Cross-validate — the agent automatically checks data across multiple sources (IFC, Excel, PDF) and reports discrepancies
• Export results — generate CSV files, summaries, and structured outputs saved to the artefacts directory

Key Findings from Proof of Concept

The proof of concept on the Karavanke project demonstrated that the agent:

• Correctly understands questions in Slovenian and maps them to the appropriate IFC terminology
• Verifies assumptions with the engineer before proceeding (10 seconds of interaction prevents minutes of incorrect analysis)
• Cross-validates results between sources (IFC, Excel, PDF) and reports inconsistencies
• Can iteratively analyze files, adapting its approach based on what it discovers

Deliverables

Deliverable	Description
Docker image	Multi-container deployment (controller + 5 workers + egress proxy)
Web interface	Chat UI with up to 5 concurrent sessions, served over HTTP
Authentication	Username/password (HTTP Basic Auth equivalent) for internal test environment
AI agent	Anthropic Max account or API key — customer chooses based on usage volume
Pre-installed tools	IfcOpenShell, openpyxl, pandas, pdfplumber, PyMuPDF, pytesseract, jq, ripgrep, sqlite3
Configurable system prompt	Mounted as a file — editable without rebuilding the Docker image
Configurable host path base	Maps container paths to user's local filesystem paths in citations
Conversation logging	All conversations persisted to a volume mount for post-analysis
Audio recording	Built-in audio capture for recording test sessions with automatic transcription
Security hardening	Read-only data, non-root user, egress firewall, capability-dropped containers

Project Data

The system provides access to the Karavanke tunnel renovation documentation:

File Type	Count	Description
PDF	969	Technical reports, drawings, specifications
IFC	276	BIM models (IFC2X3: 209, IFC4: 64)
Excel	8	Kampada registries, codebooks, inventories
DWG	182	CAD drawings
DOCX	17	Text documents
Total	~1,672	~10.7 GB