2.5M+ pages · 6 languages
Multilingual EU-law research over 2.5M+ pages, answers in 6 EU languages with full source citations, 100–200 ms retrieval.
Source: LexEU
Build vs buy RAG: what you build, what DocsGPT provides
Building a RAG assistant in-house means owning ingestion, retrieval, citations, permissions, guardrails, evaluation, interfaces, an API, observability and upgrades for as long as the product lives. DocsGPT provides those components under an MIT licence, self-hostable, with an OpenAI-compatible API, so a team can build on it instead of before it. Building from scratch is the right call when retrieval itself is the product.
Last updated: · Facts checked on
What a team must build and maintain
Ten components a production RAG assistant needs, what DocsGPT provides for each, and what stays yours.
Building in-house means writing and maintaining ingestion and OCR, chunking and retrieval, citations, permissions, guardrails, evaluation, interfaces, API, observability and upgrades. DocsGPT ships each of these under MIT; configuration, data, model choice and your own product layer remain yours.
| Build in-house | DocsGPT | Still yours to do | |
|---|---|---|---|
| Ingestion and OCR | Parsers for 30+ formats, OCR for scanned PDFs, table extraction, incremental re-indexing | Ships — Knowledge & Connectors | Choose sources; set sync schedules |
| Chunking and retrieval | Chunking strategy, hybrid (BM25 + vector) search, GraphRAG, reranking | Ships — Search & Retrieval | Pick the vector store and embedding model |
| Citations and source panel | Chunk-to-source mapping, page numbers, a UI that shows them | Ships — cited answers | — |
| Permissions, SSO, SCIM, RBAC | OIDC integration, SCIM provisioning, roles, per-source access | Ships — Admin, Security & Analytics | Connect your IdP; define roles |
| Guardrails and human approval | PII detection and redaction, prompt-injection checks, approval gates, event log | Ships — Guardrails | Choose checks and thresholds |
| Evaluation and benchmarks | Test sets, judge prompts, regression runs in CI | Ships — docsgpt-cli bench | Write the domain test set |
| UI, chat and search widgets | Chat UI, embeddable widgets, mobile layout, accessibility | Ships — widgets | Brand and embed |
| API and webhooks | OpenAI-compatible endpoint, streaming, webhooks, keys | Ships — API | Build your product on it |
| Observability and audit | OpenTelemetry traces, audit log, token and cost attribution | Ships — audit log, analytics | Route to your observability stack |
| Upgrades and security patching | Dependency updates, model API changes, CVE response, for the product's life | Ships — releases, SECURITY.md and advisories from the project | Schedule upgrades |
Component list from the DocsGPT feature set; each DocsGPT cell links to the product page for that component, and the upgrades row to the repository.
An in-house build and DocsGPT at a glance
What you get from DocsGPT versus what you write yourself.
An in-house build has no licence fee, no seat minimum and no vendor, and every component is engineering time to write and maintain. DocsGPT ships those components under MIT, self-hostable, with your choice of model and an OpenAI-compatible API; it publishes a per-seat price for Cloud only and quotes Managed, On-premises and Air-gapped flat per deployment.
| Building in-house | DocsGPT | |
|---|---|---|
| Open-source licence | Your code, your licence1 | MIT (platform and CLI)2 |
| Self-host and air-gap | Yes — wherever you deploy it1 | Yes — Self-hosted (community), On-premises and Air-gapped (zero external egress), from one seat2 |
| Model choice / BYOM | Whatever you integrate and maintain1 | Any cloud provider or local engine (Ollama, vLLM, llama.cpp and others); BYOM2 |
| Transparent pricing | No list price — no vendor; the cost is engineering time plus infrastructure1 | Yes, per posture: Cloud $20/seat/month; Managed from $2,000/month, flat; On-premises and Air-gapped from $10,000, one-time or recurring2 |
| Seat minimum / contract | None1 | None; self-hosting has no licence fee2 |
| Vendor access to data | None — no vendor1 | Yes on Cloud and Managed — Arc53 operates both, Managed as a dedicated instance in your region; No on On-premises and Air-gapped, which you operate2 |
¹ The in-house column describes what a team writes and operates itself. ·² Sources: docs.docsgpt.cloud and the public repository.
When to build
- Retrieval is your core intellectual property and the thing customers pay for.
- Scale or latency targets sit outside what a general platform is tuned for.
- A stack constraint (language, runtime, data store) that DocsGPT does not meet.
- You need one component — an embedding service, a reranker — not a platform.
The case for building on DocsGPT
- MIT licence: no per-seat cost on self-hosted deployments, no licence tax on your product.
- Self-host in your cloud or data centre; the same code runs air-gapped.
- OpenAI-compatible API, streaming API, webhooks, chat and search widgets to embed.
- Evaluation suites with
docsgpt-cli benchso retrieval and model changes are regression-tested. - Upgrades and security patches come from the project; you carry only your own layer.
docs.example.com
Ask the docsdocsgpt
How do I rotate an API key?
Open Settings › Agents, pick the agent and choose Regenerate key. The previous key stops working immediately and the new one is shown once, so update clients first 1.
→ Admin guide · API keys · p. 12
zsh — docsgpt-cli bench
6+ h → under 5 min
IP application drafting: 6+ hours to under 5 minutes per application, with lawyer review as the final step.
Source: Balt Alnoor
Cost view
In-house cost is engineering time plus infrastructure plus maintenance; DocsGPT self-hosted cost is your infrastructure plus a flat price per deployment, with the per-seat Cloud plan as the one exception.
| Build in-house | DocsGPT | |
|---|---|---|
| 100 seats, per year | Engineering time + infrastructure + maintenance | Cloud $24,000/yr · Managed from $24,000/yr · On-premises from $10,000, one-time or recurring |
| 1,000 seats, per year | Engineering time + infrastructure + maintenance | Cloud $240,000/yr · Managed from $24,000/yr · On-premises from $10,000, one-time or recurring |
| Basis | no vendor price; the cost is engineering time | Cloud $20/seat/month × seats × 12 · Managed from $2,000/month, flat · On-premises and Air-gapped from $10,000, one-time or recurring, plus your infrastructure · Cloud is the one per-seat option |
Seats matter only on DocsGPT Cloud ($20/seat/month). An in-house build costs engineering time plus infrastructure plus maintenance for the product's life. DocsGPT cost is your infrastructure plus a flat monthly fee for Managed, or a custom fixed price for On-premises and Air-gapped that covers implementation and paid production support — see pricing.
Frequently asked questions
Should I build my own RAG?
- Build when retrieval is the product you sell, or when a hard stack or latency constraint rules platforms out. Otherwise the ten components in the table above — ingestion, retrieval, citations, permissions, guardrails, evaluation, interfaces, API, observability and upgrades — are undifferentiated work that an MIT-licensed platform already provides, and the engineering is better spent on your own layer above it.
How much does it cost to build a RAG system?
- The cost is the engineering time to build and then maintain each component in the table above — ingestion and OCR, retrieval, citations, permissions, guardrails, evaluation, interfaces, API, observability and upgrades — plus infrastructure for the life of the product.
Does DocsGPT have an OpenAI-compatible API?
- Yes. DocsGPT exposes an OpenAI-compatible endpoint alongside its streaming API, webhooks and embeddable widgets, so existing clients and SDKs can point at your deployment. See API, Widgets & Channels and the OpenAI-compatible API docs.
Can I replace DocsGPT components with my own?
- Yes. Vector stores, embedding models and inference engines are configurable, and the model layer accepts any cloud provider or local engine. Which stores and engines are supported is listed on Models & BYOM and in the docs: model providers, local inference and embedding models.
Alex Tushynski
Co-Founder, Arc53
Spotted an error? Request a correction and we will fix it and update the date.
See DocsGPT on your documents
A 30-minute demo on your use case, or a free on-premises review with an engineer.