Salta al contenuto principale
Lympha technologies

Development

AI, Agents and RAG

When AI doesn't just answer, but delivers value.

A generic language model knows the world, but it doesn't know your organisation. With RAG (Retrieval-Augmented Generation) architectures we connect models to your organisation's real data — documents, knowledge bases, management systems — so answers are relevant, up to date and verifiable, with references to the sources they come from. With agents we take the next step: the AI doesn't just answer, it carries out tasks — querying systems, preparing documents, starting workflows — within rules and permissions defined by you.

It's not theory: we've built it

Our on-premise RAG platform ACME ECMS IA & RAG Integration (Private, Simple & Fast RAG) is a fully containerised stack, running on dedicated GPUs in an isolated private network: in our Private Cloud & VDC or directly in the customer's data centre. All the components are open source and the models are interchangeable (Mistral, Gemma, Qwen…): no calls to external services, no data leaving the perimeter, no vendor lock-in.

Diagram of the on-premise RAG infrastructure Inside a containerised private network: users work in Open WebUI; the platform's interface and agent query LightRAG, the RAG engine; LightRAG uses the models running on NVIDIA GPUs (Ollama for LLMs and embeddings, vLLM for re-ranking, ComfyUI for images) and stores vectors and the knowledge graph in PostgreSQL with pgvector and Apache AGE; Redis acts as the cache; pgAdmin, container monitoring and RAGAS complete day-to-day operations; persistent volumes and dedicated backups. ON-PREMISE INFRASTRUCTURE · Lympha Private Cloud/VDC or the customer's data centre · no data leaves the perimeter DEDICATED PRIVATE NETWORK — containerised stack (Docker Compose) Users corporate browser Open WebUI chat interface AI agent planner + actions LightRAG — RAG engine hybrid retrieval: vector index + knowledge graph PostgreSQL pgvector · vectors Apache AGE · knowledge graph Redis cache Persistent volumes & backups dedicated backup script, verifiable restore NVIDIA GPUs — model execution Ollama LLM & embeddings Mistral · Gemma · Qwen… vLLM — re-ranking ranks the most relevant passages ComfyUI image generation (FLUX) pgAdmin · container monitoring administration and operations RAGAS — measured quality tests and metrics in Italian
Infrastructure diagram of the platform: the real stack, containerised in a dedicated private network, exactly as we install it in our private cloud or on the customer's premises.

How it works: the integrated technologies

The logical flow has two paths — ingestion, which turns your documents into a private knowledge base, and query, which takes you from the question to the answer with sources — plus a quality measurement loop.

Logical diagram of the RAG pipeline Phase one, ingestion: documents are split up by LightRAG, which extracts entities and relationships; the embeddings computed by Ollama go into the pgvector vector index and the entities into the Apache AGE knowledge graph, forming the private knowledge base. Phase two, query: the question goes through the agent's planner, which decides between a direct answer and RAG; hybrid retrieval combines vectors and graph, vLLM re-ranking orders the passages, and the model on Ollama generates the answer with its sources, in Italian. Phase three: RAGAS generates test questions from the documents and measures the quality of the answers; Redis speeds things up with its cache. 1 · INGESTION — your documents become knowledge Documents PDF, Office, texts LightRAG chunking · embeddings (Ollama) entity and relation extraction Vector index pgvector Knowledge graph Apache AGE · entities & relations Knowledge base private, in your perimeter 2 · QUERY — from the question to the answer, with sources Question chat or API Planner (agent) does it need RAG or a direct answer? Hybrid retrieval vectors + knowledge graph (LightRAG) Re-ranking vLLM: only the most relevant passages Answer LLM on Ollama sources cited, in Italian direct answer (without RAG), if the question is not about your documents 3 · QUALITY AND CONTINUOUS IMPROVEMENT RAGAS — automatic evaluation creates test questions from your documents and scores answers Redis cache faster answers, less GPU Monitoring & backups run as a managed service
Logical diagram: the technologies integrated in the pipeline, from document ingestion to the answer with sources, with quality measured continuously.

See it in action

A simulation of the platform's real life cycle: the stack starting up, the components integrating, and a request travelling through the RAG — from the user's shell to the latent space where knowledge is indexed, through to the answer with its sources.

Demo: from docker compose up to the answer with sources

Demonstration simulation (fictitious data): the containerised stack starts up, the components integrate, then a user request arrives — the planner chooses RAG, the search travels through the latent space (vectors + graph), the re-ranking filters the passages and the local model composes the answer with its sources.

What makes our platform different

  • Not just vectors: a knowledge graph. Alongside the vector index, LightRAG builds a graph of entities and relationships (Apache AGE on PostgreSQL): answers connect the facts, rather than merely resembling the question
  • Dedicated re-ranking (vLLM): retrieved passages are re-ordered by relevance and filtered before they reach the model — less noise, more precise answers
  • A planner that decides: the agent assesses every question and chooses between a direct answer and RAG, explaining why — no pointless searches, no made-up answers
  • Quality measured, not claimed: RAGAS generates sets of test questions from your own documents and evaluates the answers with metrics, in Italian
  • Images too: the stack includes on-premise image generation (ComfyUI), with the same confidentiality as text

Governed as a service, not as an experiment

  • Data never leaves the perimeter: models, indexes and documents live in the dedicated private network — GDPR compliance by design
  • Open-source components and interchangeable models: no vendor lock-in, predictable costs (the hardware is yours or on a fixed fee, not billed per token)
  • Operations included: container monitoring, administration, persistent volumes and backups with verifiable restore — with optional oversight by our NOC
  • Tailored to you: the platform is a proven starting point; models, data sources, permissions and integrations adapt to your context

Typical use cases

  • Intelligent document search: query archives, resolutions, manuals and procedures in natural language — with the source cited
  • Internal assistants built on your corporate knowledge base: help desk, operational support, onboarding
  • Document analysis and preparation: summarising, classification, data extraction from case files and contracts
  • Agents in your processes: repetitive tasks carried out by the AI within your workflows, with human approval where needed

Who it's for: public administration, healthcare and businesses that want the benefits of generative AI without giving up confidentiality, compliance and control — or that have already tried "generic" AI and seen its limits on their own data.

Copyright © 2011–2026 Lympha Technologies S.r.l.

Conformità WCAG 2.1 livello AA