Every enterprise AI project hits the same wall. Not the model wall. Not the prompt wall. The data wall.

Companies arrive with ambitious visions: an AI advisor that knows every product in the catalog, handles complex multi-step buying journeys, and converts like the best rep on the team. Then reality arrives. Product data lives in three legacy systems. Pricing is somewhere else entirely. Documents are PDFs no one has touched since 2019. Half the catalog has missing attributes, conflicting descriptions, and no clear relationships between items.

Most teams either give up, or spend 18 months building brittle data pipelines that break the moment the catalog changes. Either way, no revenue gets generated.

At Kleio, we decided to solve the data problem properly — not as a side project, but as a core product. The result is our Knowledge Engine: the layer that transforms complex, fragmented enterprise data into something AI agents can actually reason over. In real time. In production.

‍

The Problem With “Just Use RAG”

There’s a common shortcut in enterprise AI: dump everything into a vector database and call it retrieval-augmented generation. It works fine for Q&A on a homogeneous corpus. It breaks badly for complex product catalogs.

Product reasoning in high-consideration sales isn’t a search problem — it’s a reasoning problem. A buyer looking for a seven-day tour through Southeast Asia with family-friendly excursions, or a customer searching for a used SUV under €35,000 with a panoramic roof and less than 50,000 km, or a relocating family who needs a 4-bedroom property within a specific school catchment area, or a procurement manager sourcing 500 units of an industrial product that meets a precise technical specification, or a contractor pricing a full bathroom renovation across a fragmented building materials catalog — none of these queries can be answered by semantic search alone. They require simultaneous traversal of product attributes, relational constraints, pricing tables, and enriched descriptions.

Different retrieval problems require different data structures. Semantic search needs vector embeddings. Relational constraints — pricing rules, availability windows, compatibility logic — need structured stores. Entity relationships need a graph. Real-time lookups need low-latency key-value access. Forcing everything through one technology means always making the wrong trade-off.

‍

Five Specialized Stores, One Unified Reasoning Layer

The Kleio Knowledge Engine distributes data across four purpose-built stores, each optimized for a specific access pattern. They arrive pre-configured with industry-specific ontologies — travel is not retail, automotive is not wholesale, building materials is not real estate — and are extended for each customer’s specific data model and use cases as new sources are ingested.

01  Product Store  ·  Key-Value + Vector Embeddings

The product store is the central repository for all business entities. Every object combines structured attributes, reference resources (PDFs, images, videos, specs), and vector embeddings of textual content. This combination enables rapid access by key, by indexed attribute filter, or by semantic similarity search.

The ontology adapts to each industry’s entity model:

• Travel — tour packages, accommodation, transport legs, excursions, destinations — with embeddings enabling queries like “family trip with cultural stopovers in Southeast Asia” or “romantic city break with wine experiences”
• Automotive — vehicles (make, model, year, mileage, trim, options), dealership inventory, financing offers — with embeddings for queries like “low-mileage family SUV with panoramic roof” or “sporty used coupe under €25,000”
• Real Estate — property listings (type, size, location, features), neighbourhoods, agents — with embeddings for queries like “bright apartment near schools with outdoor space” or “investment property with high rental yield in a university district”
• Wholesale — product references (SKU, category, supplier, MOQ, lead time), customer accounts, contract pricing tiers — with embeddings for queries like “indoor LED lighting with fast delivery and volume discount” or “food-grade packaging compatible with cold chain logistics”
• Building Materials — product catalogue (category, technical specs, certifications, compatible products), projects, regional stock — with embeddings for queries like “insulation panels for low-energy renovation” or “exterior cladding resistant to coastal weather”

‍

02  Pricing & Availability Store  ·  Flat Tables

Prices, availability, options, and promotions are stored in a denormalized structure where each row represents a unique combination, enabling ultra-fast queries and real-time aggregations. Each industry has its own pricing dimensions: departure dates and package tiers in travel, configuration options and finance rates in automotive, price per square metre and tenure type in real estate, contract pricing tiers and minimum order quantities in wholesale, regional stock levels and project discount grids in building materials.

Promotional logic — early-booking discounts, last-minute offers, group rates, seasonal pricing — is managed as columns in this store, eliminating join-heavy queries at inference time and ensuring agents always work with live, accurate pricing.

‍

03  Graph Knowledge Engine  ·  Relationship Traversal

The graph models relationships between entities and enables multi-level traversals that are impossible with a classical relational database. In travel, the graph connects destinations to tours, tours to excursions, excursions to themes. In automotive, it connects vehicle models to trim levels, options packages, and compatible accessories. In real estate, it links properties to neighbourhoods, schools, transport links, and comparable listings. In wholesale, it maps products to suppliers, substitutes, and account-specific contract terms. In building materials, it connects products to compatible systems, required accessories, and technical certifications.

AI agents traversing the graph can answer questions that require reasoning across the entire catalog in a single pass — for example: “Show me all tours in the South of France that include at least one wine experience and depart on a weekend in September,” or “What insulation products are compatible with this wall system, currently in stock, and eligible for the energy renovation subsidy?”

‍

04  Document Store  ·  Vector Search

This vector store contains all contractual, regulatory, and informational documentation. Customer questions are vectorized and compared semantically to automatically surface the relevant section. In travel: general terms, cancellation and refund policies, visa requirements, accessibility conditions. In automotive: warranty terms, financing agreements, return policies. In real estate: lease and purchase agreement templates, co-ownership rules, local planning regulations. In wholesale: framework agreements, supplier terms, logistics and returns policies. In building materials: product technical datasheets, installation guides, compliance certifications, and applicable subsidy or eco-label documentation.

‍

05  Customer Store  ·  Key-Value + Vector Embeddings

The Customer Store functions as the system's living memory, combining Key-Value pairs for structured precision - identifiers, budgets, preferences, loyalty status - with Vector Embeddings that capture semantic intent, enabling agents to surface offers that match what a customer means, not just what they typed. Layered on top, every past quote, converted offer, and abandoned negotiation is stored alongside the raw conversation history that surrounds it: the sessions where hesitation appeared, the objections raised, the phrases that revealed budget sensitivity.Together, these four signals give the agent a complete commercial and conversational record of each customer relationship. Rather than starting from a blank CRM entry, every new interaction is grounded in who this customer is, what they were offered, how they responded, and why they did or didn't convert - ensuring that every generated recommendation is hyper-personalized and commercially informed.

‍

‍

The Engine in Action: Query Examples Across Industries

These examples illustrate how the orchestrator routes a single customer query across multiple stores to assemble a precise, trustworthy answer — in real time, without hallucination.

TRAVEL  ·  Simple Query · Pricing Store

“What is the cheapest double room available for the Provence cycling tour departing on August 3rd?”

How the engine reasons:

Direct query on the Pricing & Availability Store: filter by tour and date, sort by ascending price. Single-store, sub-100ms response.

‍

TRAVEL  ·  Cross Query · Graph + Product Store

“Which tours include wine-tasting experiences and pass through at least one UNESCO World Heritage site?”

How the engine reasons:

The graph identifies tours tagged with wine experiences → cross-references with destinations flagged as UNESCO sites → embeddings refine by semantic similarity to surface the most relevant matches and descriptions.

‍

AUTOMOTIVE  ·  Multi-Layer Query · Graph + Pricing + Product Store

“I’m looking for a 7-seater SUV, diesel or hybrid, under €45,000, with a tow bar option and available for delivery within 4 weeks.”

How the engine reasons:

Embeddings identify 7-seater SUV models matching the powertrain preference → the graph checks which configurations include a tow bar option → the Pricing Store filters by budget and delivery window → ranked inventory with full spec sheets is returned.

‍

REAL ESTATE  ·  Contextual Query · Document + Product Store

“We’re a family of four looking to buy. What are the rules around co-ownership in this building, and are there properties with at least 3 bedrooms currently available?”

How the engine reasons:

The Document Store returns the relevant co-ownership clauses and rules → Product Store embeddings surface 3+ bedroom listings in the building or area → the graph cross-references school proximity and neighbourhood scores → a complete, structured response is assembled.

‍

AUTOMOTIVE  ·  Comparative Query · Pricing + Product + Graph

“Compare the Peugeot 5008 and the Renault Espace: differences in equipment, engine options, safety ratings, and current promotional pricing.”

How the engine reasons:

The Product Store provides full spec sheets for both models → the graph compares available trim levels, options packages, and safety certifications → the Pricing Store returns current offers and finance rates → a structured side-by-side view is generated.

‍

WHOLESALE  ·  Account-Aware Query · Pricing + Graph + Product Store

“I need 800 units of food-grade shrink wrap with a delivery lead time under 5 days. What are my options and do I qualify for a volume discount?”

How the engine reasons:

The Pricing Store filters by account tier and applies the customer’s contracted discount grid → the graph identifies substitute SKUs from approved suppliers if the primary reference is out of stock → the Product Store returns spec sheets and compliance certifications → a ranked shortlist with net pricing is returned.

‍

BUILDING MATERIALS  ·  Technical + Subsidy Query · Document + Graph + Product Store

“I’m renovating a 1970s house and need exterior wall insulation that qualifies for the energy renovation grant. What products do you recommend and what are the installation requirements?”

How the engine reasons:

The Document Store surfaces applicable subsidy conditions and eligible product criteria → Product Store embeddings identify insulation systems meeting thermal performance thresholds → the graph checks compatibility with standard 1970s wall construction and retrieves associated installation guides and required accessories → a complete recommendation with compliance documentation is assembled.

‍

Speed Is Part of the Product

We ship this in weeks, not years.

That’s possible because the ontologies are pre-built for each industry and sub-sector, the architecture is modular, and the ingestion pipeline uses AI to handle the messy reality of enterprise data — inconsistent formatting, missing fields, duplicate records, unstructured descriptions. We don’t ask customers to clean their data first. We ingest it, clean it, and augment it as part of the process.

The Knowledge Engine is fully tenant-isolated — extensible to each customer’s specific needs without cross-contamination — and designed to compound. Every new data source or use case deployed enriches the model. What costs 100% effort to build standalone costs 15–40% on Kleio.

‍

Infrastructure, Not a Feature

The Knowledge Engine isn’t a bolt-on. It’s the foundation everything else runs on.

Our orchestration layer deploys thousands of AI agents that compose, route, and collaborate to run sales conversations, ingest data, and execute complex workflows. Those agents are only as good as what they can retrieve. Give them a flat vector index and they hallucinate, miss constraints, and make recommendations that lose deals. Give them the Knowledge Engine and they reason like an expert.

This is why our customers — European market leaders including Emil Frey, Orpi, Havas Voyages, Selectour, and Showroomprivé — see 20–50% higher sales conversion rates. It’s not the model. It’s the quality of the knowledge the model is reasoning over.

Most companies treat the data layer as an infrastructure problem to solve later. We built Kleio around the conviction that it’s the only problem worth solving first.

Because in complex sales, the quality of the knowledge your AI has access to is the quality of your AI. There is no shortcut around that. We just made the right path a lot shorter.

‍

Kleio is the commercial infrastructure for complex sales — serving Travel, Real Estate, Automotive, Wholesale, and Building Materials. 100,000+ autonomous conversations per month, in production. kleio.ai

‍