All your documents stay on your infrastructure — nothing is stored on third-party servers. The AI processes everything locally, and only sends small text snippets to specialized AI models via secure API calls when generating answers.

What's an API call? When you use ChatGPT, you access one model through a subscription. Cadex AI uses multiple specialized models — one optimized for understanding text meaning, another for ranking relevance, another for writing answers. Each model is accessed via a private API key (like a password), not through a consumer subscription. This means you always get the best model for each task, and you can upgrade to newer models the day they release — no waiting.

How is it priced? API calls are billed per token — a token is roughly ¾ of a word (so "payment arrangement" = 2 tokens, a full page of text ≈ 400 tokens). Pricing is per million tokens, typically $0.10 – $15 per million tokens depending on the model. In practice: a single search query costs fractions of a cent. A Pro research report using Claude Opus + Perplexity Sonar + Exa runs ~$1.50. An Ultra debate with 7 premium agents (GPT-5.4 Pro, Claude Opus, Gemini 3.1 Pro, Grok 4.20, Grok 4.20 Multi-Agent, Mistral Large, Command R+, Qwen 3) across 5 rounds costs ~$12 — less than a single hour of consultant time for analysis that would take a team days.

Cadex AI reads every document format your company uses — including scanned PDFs with OCR (optical character recognition). It preserves tables, headings, and structure. It also extracts and classifies every image inside documents: logos, charts, photos, diagrams, and signatures.

Drop any file in and AI automatically classifies whether it's internal company data (policies, playbooks) or belongs to a specific client — and files it in the right place. If you upload the same file twice, deduplication detects it by content hash and skips the duplicate.

Documents are split into focused passages at natural boundaries — headings, paragraphs, sections. Each passage (~512 tokens) is small enough to be precise but large enough to carry full context.

Within each passage, words are broken into tokens — the smallest units AI can process. "Payment" is one token. "Arrangement" becomes two: "arrange" + "ment". This is how computers read text.

Each chunk is sent to the Embedding API, which returns a list of 1,024 numbers — a "meaning fingerprint." Similar ideas produce similar numbers, even when the words are completely different.

These numbers are stored in a vector database with an HNSW index that can search across millions of fingerprints in under 100 milliseconds.

Your question gets its own embedding. Then two independent searches run simultaneously:

Meaning search finds the 100 chunks with the closest fingerprints — catches paraphrased or differently-worded information.

Keyword search (BM25) finds chunks containing exact terms like names, numbers, and codes — catches specific identifiers meaning search might miss.

Results are merged using Reciprocal Rank Fusion, then the Reranking API reads each candidate and picks the absolute best matches. This achieves 98.7% retrieval accuracy.

The top-matched passages are sent to the Generation API as context. The AI reads them and writes a clear, natural-language answer — citing exactly which document, page, and section each fact came from.

Nothing is fabricated. Every claim is traceable to a source document. The AI only uses information from the passages you gave it — it doesn't hallucinate, guess, or make things up.

As better AI models are released — and they're released frequently — Cadex AI can switch to the latest model within hours, not months. You're never locked into yesterday's technology.

Deep Research goes beyond your knowledge base. When you need comprehensive analysis, Cadex AI searches your internal documents AND the web simultaneously, then cross-references findings and synthesizes a structured report.

You choose the depth: Standard (4 sources, fast) up to Ultra (9 sources, 4 independent AI critics that challenge every claim before the report is finalized).

Every finding includes a confidence rating and a direct link to its source — whether that's page 7 of an internal contract or a CFPB regulatory filing found online.

For decisions where there's no single right answer, Structured Debate assigns multiple independent AI agents — each powered by a different AI model — to argue opposing sides across multiple rounds. You choose the tier:

Standard 3 agents, 3 rounds — fast directional answer
Pro 4 agents, 4 rounds — deeper analysis with more perspectives
Ultra 5+ agents, 5 rounds — exhaustive, adversarial stress-testing

In each round, agents state their position with a confidence score (1-10), present evidence, and directly address each other's arguments. As rounds progress, agents concede points where evidence is convincing, and confidence scores shift in real time.

After the final round: a verdict with the recommended course of action, consensus points, specific dissents, and prioritized action items.