Sovereign Deployment

Two specialised engines working together. The search engine comprehends. The answer engine synthesises. Together they answer questions without hallucinating - because every answer is grounded in 4.8 million EU documents.

1. How Pauhu is organised

Pauhu® separates comprehension from synthesis. This is a deliberate engineering decision. We separated these two functions because combining them in a single system is what causes hallucinations. When one AI model both retrieves and generates, it invents things. When two specialised subsystems work together - one that finds evidence and one that writes from that evidence - the output is grounded in fact.

  SEARCH ENGINE                           ANSWER ENGINE
  Analytical comprehension                Grounded synthesis

  Reads your question              ───►   Reads the evidence
  Searches 4.8M documents                 Generates a grounded answer
  Classifies by topic and domain          Renders in 24 languages
  Applies regulatory rules                Presents in your browser
  Guards data integrity                   Produces the response

              ═══════════════════════
              ║   CONTROLLED  DATA  ║
              ║       FLOW          ║
              ║ (verified evidence  ║
              ║   only crosses)     ║
              ═══════════════════════

  INDEX        ◄── 4.8M+ documents + 2.4M+ terms (shared)   ──►
  SYNC         ◄── automated sync keeps data fresh           ──►
  GATEWAY      ◄── single entry point, request validation    ──►

The gateway

A single entry point receives every query, validates it, and routes it to the correct engine. The gateway decides which data sources, language models, and domain specialists are relevant to your question - before either engine does any work.

2. Search engine - comprehension

The search engine is where understanding happens. When you ask a question, this is the side that reads it, determines what you need, and finds the relevant evidence from nearly five million documents.

3. Answer engine - synthesis

The answer engine takes the evidence found by the search engine and produces the output you see. It reads multiple passages simultaneously and writes a single coherent answer - always grounded in the source documents, never fabricated.

4. The data flow between them

A dedicated data flow carries verified evidence from the search engine to the answer engine. This is not a single API call - it is a structured pipeline that ensures:

• Only verified evidence crosses. The search engine's quality checks must pass before data reaches the answer engine. No unverified claims, no unchecked sources.
• Every crossing is auditable. Each passage that moves from comprehension to synthesis is logged with a SHA-256 integrity hash. Your compliance team can reconstruct exactly what evidence informed each answer.
• The pipeline is directional. Evidence flows one way only. The answer engine cannot reach back into raw data - it can only work with what the search engine has validated and delivered.

5. How it stays current

EU institutions publish new legislation, court rulings, procurement notices, and regulatory updates continuously. Pauhu's automated synchronisation keeps the data current:

• Daily - all 20 EU institutional sources and 28 national law databases are synced every processing cycle
• Weekly - statistical updates, chemical substance registrations, institutional directory changes

When new data arrives, it flows through the same two-engine pipeline: the search engine indexes, classifies, and validates it. Only then does it become available to the answer engine for synthesis. This means the system never serves an answer based on data it hasn't verified.

6. The sovereign deployment

Everything described above - both engines, the data flow, the index, the automated sync - can run entirely on your hardware. This is the sovereign deployment: the same system, the same models, the same data, with no cloud dependency and no data leaving your premises.

What changes in a sovereign deployment

Hardware requirements

Delivery and installation

Three delivery methods: encrypted download via SFTP, physical media for classified environments, or push to your private container registry. Installation is a single command:

# Load and run the sovereign deployment
docker load -i pauhu-sovereign.tar.gz
docker run -d --name pauhu-sovereign --restart unless-stopped \
  -e PAUHU_SOVEREIGN=true -p 3000:3000 pauhu/sovereign:latest

# Verify: search for EU AI Act
curl http://localhost:3000/v1/search?q=artificial+intelligence+regulation

No configuration files, no API keys, no cloud accounts, no database setup. The container includes both engines, the data flow pipeline, and the complete data snapshot.

7. What data is included

The sovereign deployment ships with a complete snapshot of all 20 EU institutional data sources - the same data that feeds the cloud version's search engine:

Total: 4.8M+ documents plus 2.4M+ terminology entries in 24 languages. This is the shared index that both engines draw from.

8. What it guarantees

9. Supply chain sovereignty

Most AI systems depend on a chain of external providers: cloud compute, proprietary APIs, third-party model hosting, and centralised inference services. Remove any link in the chain and the system stops working. This is a single point of failure - or multiple single points of failure.

What supply chain sovereignty means

• No API dependency: Pauhu does not call external AI APIs. The models are optimized files that execute locally - in the browser via browser-native processing, or on the server via a lightweight runtime. If every cloud provider went offline simultaneously, Pauhu would still work.
• No model hosting dependency: The models ship with the container or are downloaded once to the browser cache. No ongoing model-as-a-service subscription. No inference-per-token billing.
• No data dependency: The 4.8 million EU documents are included. You do not need to query an external database. The data is yours, on your volume.
• No vendor lock-in: Open model format (ISO/IEC 17203). Docker containers are OCI-compliant. The REST API follows OpenAPI 3.1. Every component uses open standards.

The geopolitical dimension

Government agencies increasingly recognise that depending on foreign-controlled AI infrastructure creates a strategic vulnerability. Executive orders, sanctions, licensing changes, or corporate acquisitions can cut off access to critical AI services overnight. Pauhu's sovereign deployment eliminates this risk: the entire system - models, data, inference - is under your control, on your soil, subject to your laws.

10. Adaptive model loading

Pauhu adapts to the hardware it runs on. Not every deployment has a GPU server with 32 GB of RAM. A civil servant’s laptop, a ministry’s standard-issue workstation, a dedicated inference server - the same architecture works on all of them, at different performance levels.

Three tiers

Why 300 MB matters

The global semiconductor supply chain is under sustained pressure. Memory prices fluctuate, procurement cycles lengthen, and government IT budgets rarely include high-end GPU servers. Pauhu’s grounded generation model fits in 300 MB of DRAM - less than a typical browser tab. This is not a limitation; it is a design decision. A model that fits in commodity hardware is a model that every government agency can deploy without special procurement.

Progressive download

Models are loaded in priority order, not all at once:

1. Search models first - paragraph retrieval is available within seconds of start
2. Grounded generation second - grounded answers become available in 10–30 seconds
3. Translation on demand - only the language pairs you use are loaded. Finnish-English loads on first Finnish query, not at startup

11. Double Anti-Hallucination

Most AI systems rely on a single layer of defence against hallucination: either they check the output after generation, or they constrain the input. Pauhu uses both - simultaneously.

The result: every claim in a Pauhu answer traces back to a specific paragraph in a verified EU document. If the system cannot ground a statement, it says so - rather than inventing a plausible-sounding answer.

12. For procurement officers

Why government agencies choose sovereign deployment

• Data sovereignty: Your queries and results never leave your premises. No cloud processing, no data transfer to third countries, no dependency on foreign infrastructure.
• Classified environments: Works in air-gapped networks, SCIFs, and restricted environments where internet access is not available or permitted.
• GDPR Article 44: No third-country data transfers. All processing happens within your jurisdiction.
• EU AI Act Article 53: Full training data transparency. Every model includes a published summary of its training data. See the AI transparency disclosure.
• Grounded by design: The dual-engine architecture ensures every answer is traceable to verified EU institutional documents. The system cannot hallucinate because the answer engine only works with evidence the search engine has validated.
• No vendor lock-in: Standard Docker container, standard REST API, open-format models. If you stop using Pauhu, your data and audit trail remain on your hardware.

Tender-ready specifications

For public procurement (CPV code 72000000 - IT services):

• On-premises deployment with zero cloud dependency
• EU-origin software (Pauhu Ltd, Helsinki, Finland, Y-tunnus 3425757-6)
• All training data sourced from EU institutional open data
• Zone-based security architecture
• WCAG 2.1 AA compliant web interface
• REST API with OpenAPI specification
• Docker container (OCI-compliant)
• Open model format (ISO/IEC 17203-compliant runtime)
• 24 EU official languages supported

Contract model

13. Frequently asked questions

Why two engines instead of one AI model?

Single-model systems generate text from statistical patterns. They can produce fluent, confident answers that are completely wrong. By separating comprehension from synthesis into two specialised engines, we ensure that the generation side can only work with evidence the comprehension side has verified. The result: grounded answers with citations, not plausible-sounding fabrications.

Does it really work offline?

Yes. After installation, you can disconnect the server from the network entirely. Both engines, the data flow between them, and all 4.8 million documents run locally. We encourage you to verify this with your network monitoring tools.

How fresh is the data?

The cloud version receives daily updates. The sovereign deployment contains a snapshot at the time of delivery. Data updates are delivered periodically via secure transfer - typically monthly or quarterly. The update process is a single command.

What hardware do we need?

A standard server with 16 GB RAM and 100 GB storage. No GPU required - all models run on CPU. A GPU speeds up inference but is not necessary. See Section 6 for full requirements.

Can we run it in a VM?

Yes. Docker runs in any virtualisation environment: VMware, Hyper-V, KVM, or bare metal. The container has no hardware-specific dependencies.

Is the source code available?

The sovereign deployment is provided as a container image. Source code review is available under NDA for government customers. Contact sales@pauhu.eu.

Can we integrate it with our existing systems?

The sovereign deployment exposes a standard REST API. Any system that can make HTTP requests can use it. API documentation is included in the container.

What is the licensing model?

Annual subscription per deployment. Volume discounts available for multiple installations. Contact sales@pauhu.eu for pricing.

Related documentation

• Grounded Generation Architecture - technical deep-dive for your engineering team
• The Guide vs. the Encyclopedia - why Pauhu exists
• Government Procurement Demo - 10-step walkthrough using Finnish government procurement
• MCP Sovereign Mode - developer reference for IDE integration
• Compass Search - how the 20 data sources are indexed