10xLab Accelerates Enterprise RAG Deployment with Inferdat on AWS

10xLab's fully on-premise RAG architecture, while well-suited to clients with strict data sovereignty mandates, required significant local GPU infrastructure to run its model pipeline, including a 70B parameter reasoning model. This created a ceiling: prospects who wanted to move faster or avoid managing GPU infrastructure had no path into 10xLab's product. 10xLab needed the same knowledge assistant to also run natively on AWS, without forking the codebase or building a second product.

Inferdat kept RAGflow as the constant document intelligence layer across both deployment modes, handling parsing, chunking, and retrieval orchestration identically regardless of where the system runs. For the AWS deployment, Inferdat replaced the local Ollama-served model layer with Amazon Bedrock, providing managed foundation models for query routing, retrieval, reranking, and deep inference. The same hybrid public/private retrieval pattern was preserved: a PII classifier and query router determine what context is safe to send to public sources, private retrieval runs against Aurora PostgreSQL within the client's AWS account, and Bedrock handles inference on the fused context. Inferdat built the customer-facing knowledge assistant UI/UX once, with a deployment-agnostic experience that works identically whether the backend is on-premise or AWS.

The key architectural decision was isolating the model-serving layer as the only component that changes between deployment modes. RAGflow's document parsing and chunking, the PII boundary classifier, the hybrid public/private retrieval pattern, and the knowledge assistant UI are identical across both paths. In the on-premise configuration, Ollama serves open-weight models (Mistral 7B, Qwen 2.5-32B, DeepSeek-R1 70B) on local hardware. In the AWS configuration, Amazon Bedrock serves the equivalent pipeline roles, with on-demand pricing eliminating the idle GPU cost the on-premise path requires. This gives 10xLab a consistent answer to infrastructure questions during sales conversations, with the choice coming down to the client's governance posture and budget rather than feature availability.

10xLab now offers a single knowledge assistant product with two deployment paths, widening their addressable market to include clients who don't want to manage GPU infrastructure. The AWS deployment delivers query routing and retrieval in under two seconds, with Bedrock handling deep reasoning across 128K context windows for technical document analysis. Because RAGflow remains the constant ingestion layer, indexed client document sets are portable between deployment modes if a client's infrastructure requirements change.