Maya HTT launches on-premises RAG assistant as forecasters project $43.75bn market

M-Bot runs inside customer infrastructure. Maya HTT says the assistant keeps sensitive engineering data on the buyer's own systems.

Buyers get two promises. Sensitive engineering data stays on their own infrastructure, and costs are fixed up front rather than metered per token. That second point matters because RAG projects often copy files into a separate vector database. That process strips permissions and leaves duplicates behind. What nobody has shown publicly is whether an on-premises assistant answers complex technical questions accurately at scale. For engineering teams, the appeal is clear. They can query internal documents without sending them to an outside model. The open question is accuracy at scale. That gap leaves buyers to test the system on their own documents.

M-Bot connects to internal engineering, manufacturing and industrial operations knowledge. It returns answers in seconds, according to the company.

The retrieval-augmented generation architecture is built for technical content. That includes text, images, tables and graphs that Maya HTT says conventional LLMs cannot parse. It links to Teamcenter and other PLM environments, plus SharePoint and Confluence. It traces each answer back to its source document. Maya HTT claims retrieval up to 60 times faster than manual search: a five-minute hunt becomes roughly five seconds. Those numbers come from the vendor, not from independent testing. The company also says M-Bot carries no LLM token metering costs, so customers know the bill before deployment.

Remi Duquette is Maya HTT's industrial AI vice president. He said engineering organizations hold large amounts of technical knowledge. They struggle to get the right information to the right person in time to influence decisions.

The target industries are aerospace, defense and manufacturing. Intellectual property rules push buyers toward systems that run on-premises or inside customer-controlled infrastructure.

The market behind the launch

A ResearchAndMarkets report published on 30 October put the global RAG platforms market at $4.00 billion in 2025 and $5.52 billion in 2026. It projected $43.75 billion by 2036, a 23.0% compound annual growth rate across the decade. The scope is broad. It counts vector databases and search engines, embedding and reranking models, orchestration frameworks, ingestion and parsing tools, connectors, evaluation and guardrail platforms, cloud-managed RAG services and enterprise knowledge assistants. It excludes foundation model usage charges and general-purpose databases without retrieval features. That definition matters for buyers comparing vendors.

Adoption is heaviest in law, finance and healthcare, where outputs need citations, the report says. AI agents are pulling demand too. Multi-step workflows need continuous access to data, applications and tools. The report points to Elastic as evidence of demand. Elastic reported more than 2,200 Elastic Cloud customers running generative AI workloads in the first quarter of fiscal 2026. It also reported an 18% rise in third-quarter revenue to $450 million.

It also treats enterprise knowledge assistants as part of the market. That is the category M-Bot sits in, not a separate tooling layer.

Accuracy and security remain the hard part

Accuracy is where the sales pitch meets evidence. A preregistered Stanford study found that RAG-based legal research products from LexisNexis and Thomson Reuters hallucinated on 17% to 33% of queries. GPT-4 used independently hallucinated on 43% in the same study. Retrieval helped. It did not close the gap. The study logged fabricated authorities, incomplete answers and misread case law. The range runs from 17% to 33%, not a single figure. Those results set a bar for any vendor selling retrieval as a fix for hallucination.

Security has its own track record. Microsoft patched EchoLeak, CVE-2025-32711, in June 2025. It was a zero-click flaw rated 9.3 on the CVSS scale. Malicious content retrieved from an email could manipulate Microsoft 365 Copilot and expose sensitive information. The report says enterprises are responding by prioritising permission-aware retrieval, prompt injection detection, data loss prevention, provenance and audit trails. Microsoft now tests Copilot against synthetic companies. Its Industry Evaluation Data Generation Platform builds simulated enterprise environments from public sources including O*NET occupational data, job postings and SEC filings. No customer data enters the loop. Microsoft says the work is meant to establish the first standardized evaluation baseline for Copilot.

RAG or fine-tuning: the architecture choice

Every enterprise RAG purchase runs into the same comparison. RAG changes the information a model gets at answer time; fine-tuning changes how the model responds by adapting its learned behaviour. The first fits documents and policies that change often. The second fits consistent task execution and fixed output formats. Parameter-efficient methods such as LoRA reduce the compute by training a smaller set of additional parameters rather than the entire model. Fine-tuning still does not give a model reliable access to the latest company information. It should not be treated as a guaranteed way to remove hallucinations.

Data handling decides much of the rest. The Dell AI Data Platform lets models query files on-premises without duplicate copies or altered file access controls. IT Pro described it on 29 October. It pairs NVIDIA AI Enterprise software with GPU acceleration. Copying files into a separate vector store creates shadow data, spends bandwidth and strips original permissions, IT Pro reported.

What happens next

Maya HTT says M-Bot is available for demonstration and enterprise deployment.

The ResearchAndMarkets forecast runs to 2036, when it puts the market at $43.75 billion, up from $5.52 billion in 2026. Microsoft describes its synthetic evaluation platform as the start of a standardized baseline for Copilot. It is not a finished benchmark.