Do not index
Do not index
One of the most common questions we see from consultants and coaches right now is some version of this: they have years of frameworks, course content and call recordings, and they want to turn all of it into a knowledge base their own AI agents can query, so answers sound like them and each client gets a separate memory. The real question underneath that is whether to build it themselves, pay for a managed service, or use a platform that already does the whole job.
Most online comparisons answer this badly, because they conflate two different systems. There is the expert's shared knowledge (frameworks, course lessons, books, approved call excerpts) and there is each client's private memory (goals, decisions, commitments, progress). Finding a relevant passage in a transcript is not the same as remembering that a client changed their mind about a deadline three weeks ago. A knowledge base alone does not produce a reliable memory system.
Is a DIY Stack With n8n, Supabase and Claude Worth the Engineering Time?
The common do-it-yourself stack looks like this: n8n for ingestion and automation, Supabase with the pgvector extension for clients, conversations and structured memory, and the Claude API for generation. Anthropic does not currently offer its own embeddings model, so a separate provider such as Voyage AI is needed for the vector search layer.
Software costs are genuinely low. Supabase Pro runs $25 a month, including some compute credit. n8n is free to self-host, or roughly 20 euros a month on its Cloud Starter plan for a few thousand workflow executions. Claude Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output tokens, which works out to around $24 a month for 1,000 typical answers. Transcribing 500 hours of call recordings costs roughly $135 at current OpenAI transcription pricing. A realistic early software budget sits between $75 and $300 a month before maintenance labor.
The part every cheap-RAG-stack post leaves out is build and maintenance time. We've covered the broader build vs. buy decision for an AI coaching platform before; this goes one level deeper into the DIY stack itself. A real production version, with authentication, content ingestion, chunking, memory extraction, a chat interface and security testing, runs 120 to 300 hours. After launch, plan on 5 to 15 technical hours a month, or $500 to $3,000 a month if an outside developer maintains it. Client isolation has to be enforced at the database level with row-level security, with the tenant ID always derived from an authenticated session and never trusted from client input.
DIY makes sense when the AI becomes a core part of a paid product, when custom integrations with a CRM or course platform are required, or when full control over data retention and deletion matters strategically. It makes much less sense as a way to save a $100 to $500 monthly subscription, since a managed service that saves eight technical hours a month is usually saving well over $1,000 in labor.
What Do Managed RAG-as-a-Service Platforms Actually Cost?
Dify Cloud sits closer to a full application builder than a pure retrieval service. Its Professional plan is $59 per workspace per month for 500 knowledge documents and 5 GB of storage, and its Team plan is $159 a month for 1,000 documents and 20 GB, on top of model usage. A small production setup lands somewhere between $85 and $260 a month. Dify still leaves the tenant model, correction process and deletion logic up to the builder.
Ragie focuses specifically on ingestion and retrieval: parsing, transcription, chunking, hybrid search and reranking across documents, audio and video. Its Starter plan is $100 a month for 10,000 processed pages, Pro is $500 a month for 60,000 pages, and extra pages cost 2 to 5 cents each, with audio processed at roughly two-thirds of a cent per minute. Ragie's documentation explicitly recommends a "partition" keyed to a user or organization ID for multi-tenant isolation, a meaningfully stronger guarantee than a loose metadata filter. Even so, the surrounding application (login, structured client memory, the chat interface itself) still has to be built around it.
Why a Custom GPT or Claude Project Is a Prototype, Not a Product
A Custom GPT can hold up to 20 files, does not carry memory between conversations, and lives inside ChatGPT rather than offering an API for an external product. Claude Projects go further: paid plans apply retrieval automatically as a project's knowledge grows, up to roughly ten times normal capacity, and Claude Pro is $20 a month while Claude Team runs $25 per seat a month annually with a five-seat minimum. One project per client can work as a manual, high-touch workaround, but it is an administrative process, not an application architecture with an API, lifecycle controls or real tenant isolation.
Both are genuinely useful for validation: upload a curated sample of material, run 50 to 100 real questions through it, and see whether the answers actually sound like the expert before building or buying anything further. Once a knowledge base is actually live, keeping it current is its own ongoing problem.
When a Dedicated Expert-Brain Platform Beats Building Your Own
A newer category of platform builds the per-user memory problem directly into its API. UseBrian explicitly separates a shared "knowledge base" from "memory," which it scopes to a specific user and keeps keyed to a stable external user ID; its hosted pricing runs $20 to $200 a month depending on usage credits. Delphi's public pricing lists a Free tier, a $79-a-month Builder plan and a $299-a-month Scaler plan, with API access only available through its custom-priced Immortal tier, so its terms for an API product with many individually remembered clients are not transparent. Personify prices a Pro plan at $29 to $39 a month and an Agency plan at $149 a month for three client workspaces, then $29 for each additional one, which is fast for a branded coaching clone, but saved conversation history is not the same as a verified, correctable memory system.
For an expert whose actual goal is a sellable subscription rather than general-purpose developer infrastructure, platforms like BuddyPro fold the same four pieces, ingestion, retrieval, per-subscriber memory and delivery, into one no-code product. An expert uploads existing books, courses, call recordings and frameworks, the AI analyzes and maps them into specialized knowledge roles, and every subscriber gets their own ongoing memory that the AI never mixes with another subscriber's history. That is launched directly to the expert's existing audience as a premium, recurring subscription in the $1-2K-a-year range, with the expert keeping 100% of the profit and typical margins of 75-85% after AI usage costs of roughly $15 to $30 per subscriber a month. David Říha has compared the tools available for building an AI clone from existing knowledge in more detail.
Option | Build time | Monthly cost | Client isolation |
DIY: n8n + Supabase/pgvector + Claude | 120-300 hrs for production | $75-$300 + maintenance labor | Strong, if row-level security is enforced correctly |
Dify Cloud | 20-80 hrs | $85-$260 | Builder designs the tenant model |
Ragie | 30-120 hrs | $100-$500+ | Strong, partition-based |
Custom GPT / Claude Project | 4-20 hrs | $20-$150+ | Manual, not application-level |
Expert-brain platform (UseBrian, Delphi, Personify, BuddyPro) | 10-60 hrs, or none on a self-serve platform | $20-$500+, or a $197/mo license on BuddyPro | Built around stable user identities |
The Real Decision
Across every option, the software bill is rarely the deciding factor. Model and hosting costs are usually small; engineering time and ongoing maintenance are what actually separate a weekend prototype from something a paying audience can rely on. A Custom GPT or Claude Project is the right starting point to test whether an expert's content produces good answers at all. Supabase plus Ragie plus Claude, or Dify, fit a team that wants to own a custom application and has the engineering capacity to maintain it.
For an expert whose goal is a monetized AI product sold directly to their own audience rather than a piece of internal tooling, going straight to a finished expert-brain platform, built on BuddyPro or a comparable alternative, skips the entire build-vs-buy decision tree and the months of engineering it would otherwise take.
Related reading:
Written by