RAG as a service for agencies: one knowledge base per client and a single API
Published on 2026-11-16 by ArduBot
RAG SaaS product model for agencies: client onboarding, limits, permissions, billing, metrics, and how to validate real demand.
An agency sells five chatbots and ends up maintaining five different systems for uploading PDFs, creating embeddings, and correcting responses. That repeated work could become a reusable platform. The opportunity is not selling “AI that knows everything,” but reliable document retrieval as a service.
What an agency would really buy
A service with which it can create a knowledge base per client, upload sources, keep them synchronized, and query documents from any chatbot or automation. The agency wants to deliver the project without managing a different indexing engine each time.
A first proposal could include: separate bases per company, PDF and URL import, visible indexing process, search API with citations, source deletion, per-project keys, and usage limits. Advanced OCR, private connectors, and integrated generation could be evaluated as later features based on real demand.
The product must not promise magic
A RAG can fail due to poor extraction, contradictory documents, missing information, or incorrect permissions. The service should show what was processed, what failed, and what evidence supports each result.
Also, it is not advisable to claim that a RAG node replaces a CRM API or an inventory database. Recent transactional data should be queried in authorized systems.
Onboarding in a commercial integration
1. The agency creates the client space and configures identity and permissions. 2. Uploads a test set of PDFs or URLs and verifies extraction and versions. 3. Runs 20-30 questions with expected sources before connecting the real bot. 4. Integrates the API from n8n, its backend, or a query MCP agent. 5. Monitors response rates with evidence, errors, and costs per client.
The key is delivering a repeatable system with clear liability limits, not simply a convincing chatbot in a demo.
How it could be charged
A flat fee per base or company simplifies predictability, while charging per processed pages, storage, or searches aligns pricing with consumption. A hybrid model could separate a monthly minimum from additional volumes. We do not publish rates because we are still validating the product and its costs.
Before setting them, it is advisable to simulate small agencies with ten bases and larger companies with thousands of documents. Support, synchronization, deletion, and security generate costs even if there are few queries.
How to know if there is a real market
Informational Google searches only demonstrate curiosity. The most useful commercial signal is that an agency delivers real documents, describes its integration, and agrees to participate in a test with limits. Later, willingness to pay, retention, and time saved compared to developing the system from scratch can be measured.
For the SEO experiment, record per article impressions, clicks, visits to the page of interest, and qualified requests. Classify requests by origin (PDF, web, Drive), volume, channel (API, WhatsApp, n8n, MCP) and urgency level. Do not count as a lead anyone who reads a tutorial.
ArduMaker's hypothesis
We are evaluating creating an automatic RAG node with API to retrieve information from uploaded sources, aimed at developers, agencies, and SaaS. It is not an available product nor a promised release date. We want to verify if the problem of maintaining documentary knowledge justifies a business unit.
If it would be useful for a client or for several projects, describe: what files you need to upload, how many, where they are updated, and what system would query the API. You can initiate contact from ArduMaker.
References and context: AWS Knowledge Bases · Microsoft RAG · Automatic RAG: ArduMaker guide.