LIRA · Libreria di Informazioni e Risorse Aziendali
LIRA — The AI that knows your documents.
Precise, verified answers with source citations — directly from your documents. No data sent outside: everything on-premise.
Built on your documents. Not on the internet.
Generic AIs are powerful but have a structural limitation: they don't know your company. They don't know what's in your contracts, your technical manuals, the protocols you've been using for twenty years. So they generalize. They make up answers. You need facts, not fantasy.
LIRA is not a pre-packaged platform to adapt to you. It's the opposite: the platform molds itself to you. The results are qualitatively superior to any out-of-the-box solution, for a simple reason: they start from the internet, we start from your documents.
How it works
Connect your sources
LIRA integrates with your systems via native connectors (filesystem, cloud storage, database, FTP, SharePoint) and custom connectors for proprietary databases. It automatically monitors your sources: every new document and every change is detected and processed without manual intervention.
Auto-index
Every added or modified document is indexed without manual operations. Semantic embeddings, keywords, images: everything on-premise, always up to date.
Ask and receive
Ask a question in natural language. LIRA retrieves the relevant passages and generates an answer with precise source citation.
Four pillars that make the difference
Total data sovereignty
LIRA is 100% on-premise by design. No data traveling to external services, no third-party API keys, no 'trust us' promises.
- Pipeline, models, vector store, LLM — everything runs within your perimeter
- GDPR-compliant, auditable, ready for even the most restrictive security policies
- Hybrid deployment (part on-prem, part cloud) or 100% cloud when needed — without redesigning
Truly scalable
Containerized, modular architecture — starts from a single machine and scales to clusters on distributed infrastructure.
- 10,000 pages of documentation or 10 million? The same platform, sized accordingly
- Embedding, reranking, LLM serving, retrieval — each component scales independently
Tailored to you
The platform molds itself to you, not the other way around. Every parameter is configurable and tuned to your context.
- Domain-appropriate embedding model (directional, technical, multilingual)
- LLM chosen based on the quality/cost/latency trade-off you need
- Chunking tuned to document type (contracts vs manuals vs reports)
- Retrieval and reranking thresholds optimized on your actual data
- Connectors for filesystem and SMB shares; SharePoint, Documentum, ERP on request
- Prompts and UI branded as core to your brand
Multimodal and verifiable
Indexes text, tables and images. Every answer — every single answer — cites the source passage.
- Text, tables, images, diagrams and plant schematics — everything enters the knowledge base
- Images receive automatically generated captions from vision models
- Full source transparency: every answer cites the source passage, document and page
Behind the scenes: the two-phase architecture
The product's core is a RAG system: given the user's question, it retrieves relevant passages from your documents and provides them to the LLM as context to generate an answer grounded in those passages — not in the model's memory.
RAG-U — the retrieval engine
RAG-U is the ingestion, indexing and retrieval pipeline that feeds the LLM with context extracted from your documents.
- Ingestion — reads your documents (.docx, PDF, mixed formats), extracts text, tables and images, splits into meaningful chunks.
- Indexing — transforms each chunk into semantic vectors + BM25 keyword index, inside an on-premise vector store.
- Retrieval — given the user's question, retrieves the most relevant passages by combining semantic and keyword search, reranks them, reconstructs section context.
- Image indexing — a parallel worker generates text captions for images and indexes them semantically alongside the text.
LIRA Assistant — the conversational interface
On top of RAG-U runs LIRA Assistant, the interface users actually interact with: a chat that responds in natural language, cites sources, and can be branded and integrated into your workflows.
- OpenAI-compatible API → integrates with any tool or framework that already uses the OpenAI Chat Completions API
- MCP server → exposes retrieval as a tool to MCP clients and external AI agents
In summary: RAG-U retrieves the context, LIRA Assistant uses it to respond.
Where it makes a difference, concretely
| Scenario | Generic online chatbot | LIRA |
|---|---|---|
| The packaging line is down: which Modbus registers should I read to check if the fault is on belt 3? | Explains in theory how Modbus works | Read register 40120 (belt 3 status): if bit 2=0 the motor is in fault. PLC manual LC3 rev.7, § 4.3 |
| The supplier delivered 18 days late. How much is the penalty? | Quotes the Civil Code, not your contract | 1.5% per day beyond 10 days = 12% of the amount. Art. 8.2, supply contract 03/15/2025 |
| Which PPE is required to access the mixing department according to our DVR? | Lists generic PPE required by law | Helmet, cut-resistant gloves, S3 footwear and self-contained breathing apparatus for ATEX zone 2. DVR rev.12, p. 47 |
| Find the CT scan of patient 4821 and the related report from June 12, 2024 | Cannot access your medical archives | 2 images + report: 'Upper abdomen CT — pat. 4821' dated 06/12/2024, signed by the radiologist. From the corporate RIS system |
Who it's for
- Production and plant directors — instant access to technical manuals, operating procedures and maintenance history, without hours of archive searching
- IT managers and CTOs — on-premise deployment, no third-party data sharing, OpenAI-compatible API, MCP server for integration with existing tools
- Legal and compliance departments — answers with precise passage citations, contract clause verification in seconds, complete audit trail
- R&D and knowledge management — always-accessible corporate memory: patents, technical reports, product specifications, historical research
- Regulated sectors — healthcare, finance, defense, energy: data that cannot leave the corporate perimeter
- Every company that tried an off-the-shelf chatbot and stopped believing in it
What you get
- RAG-U, the retrieval-augmented generation pipeline in on-premise or hybrid deployment
- LIRA Assistant, the conversational interface your users interact with, brandable
- OpenAI-compatible API → integrates with any tool or framework that already uses the OpenAI Chat Completions API
- MCP server → exposes retrieval as a tool to MCP clients and external AI agents (orchestration, multi-step, automation)
- Connectors for continuous content ingestion (filesystem, SMB shares, others on request)
- Multimodal indexing of text, tables and images
- Team training and dedicated engineering support
Stop searching. Ask LIRA.
Your enterprise AI.
Free assessment and demo on your sample
Send me a sample of your documentation: I'll show you LIRA on your documents, not on a generic benchmark.