Build vs buy: AI documentation assistant
Last reviewed:
Many engineering teams look at the OpenAI Responses API, LangChain or LlamaIndex and think 'we could build this ourselves in a week'. A prototype does take days. A production assistant — one that stays in sync with your docs, refuses to guess, works in Slack and your helpdesk, and shows you what it couldn't answer — takes months, and someone has to own it after. This page breaks down what each path involves. The time and cost figures are rough estimates, not measurements.
Choose to build if the assistant is your product, or your knowledge doesn't fit docs, wikis and tickets, and you can staff it for years. Choose Kelu if you want cited answers from your docs on your site, in Slack and Teams, over MCP and in your helpdesk now — with an abstention gate, analytics, and an API and SDKs for the parts you want to make your own.
At a glance
Time and cost figures for the build side are rough, order-of-magnitude estimates for a typical docs assistant, not measurements. Your numbers will differ.
What 'build' means today
There are three common starting points. None of them is a docs assistant on its own. OpenAI Responses API. The replacement for the Assistants API, which OpenAI shut down on August 26, 2026. Its hosted file search tool runs semantic and keyword search over files you upload to vector stores, merges the two with reciprocal rank fusion, and returns file citations. ChatKit gives you an embeddable chat UI. Getting your docs into vector stores, and keeping them current, is up to you. LangChain. An open-source framework for building agents, with LangGraph for orchestration and LangSmith for tracing and evaluation. You pick the document loaders, the vector database, the retriever and the model, and wire them together. LlamaIndex. An open-source framework for RAG and agents, from the company behind LlamaParse. LlamaParse handles hard documents such as scanned PDFs and offers managed indexing. You still build the surfaces, the sync and the answer checks. All three are good tools. The work they leave is the same: connectors, sync, a check that refuses to answer, the widget, bots, helpdesk integrations and analytics.
When building is the right call
Two situations genuinely favour building. The assistant is your product. If the AI you build is the product — a coding agent, an on-brand character, a domain-specific reasoning system — the flexibility of a framework is worth the engineering. A packaged docs assistant won't be your differentiator. You have unusual retrieval needs. If your knowledge doesn't fit 'docs + wikis + tickets + code' — structured claims data, medical records with strict lineage, a proprietary knowledge graph — off-the-shelf retrieval may not model your data well. That's a real reason to build. Outside those two, building a docs assistant usually starts as 'a two-week job' and turns into a standing commitment.
When buying is the right call
Most other cases. If your problem is 'we have docs, wikis and tickets, and we want an assistant that answers from them with citations, in our widget, Slack, Teams, helpdesk and AI tools, without making things up', that is exactly what Kelu and its peers are built for. The time you'd spend building the platform is time not spent on your product. The buy path also gives you something in-house builds tend to postpone: abstention from day one. Kelu checks whether the retrieved docs cover the question before it answers, and says it doesn't know when they don't — no answer and no citations. A prompt that says 'only answer from the context' looks like it does this, until a real user asks something your docs don't cover.
What the hybrid looks like
Some teams end up in the middle: buy the platform for most of the work, and build the part that is specific to them. Kelu's public REST API, hosted widget, MCP server and React and JavaScript SDKs are made for this. You get the packaged connectors, retrieval and analytics, and build your own experiences (mobile, in-product help, agent workflows) on top. That's usually the answer when 'buy' feels like it leaves something on the table. You're not choosing between packaged and custom; you're putting custom on top of packaged.
Got questions?
Frequently asked questions
Can't LLMs just answer from our docs without RAG now?
Not reliably. Context windows are large, but a model still can't answer 'what changed in v4.2' without retrieving your changelog, and pasting every doc into every question is slow and expensive. Retrieval is what makes the answer current and citable. What has changed is that MCP lets AI tools such as Cursor, Claude and ChatGPT retrieve from your docs directly, which is why Kelu runs an MCP server per knowledge base.
Is the OpenAI Assistants API still an option?
No. OpenAI deprecated it in August 2025 and removed it on August 26, 2026. The replacement is the Responses API with the Conversations API, and file search and vector stores carry over. If you built on Assistants, you have already had to migrate once — a good example of the upkeep a build carries.
How much does Kelu actually cost at scale?
Pro is $99 a month, or $79 a month billed yearly (1,000 questions a month, 5 knowledge bases, 10 team members). Enterprise is custom, with unlimited knowledge bases, pages, questions and seats, SSO, a white-label widget, source groups and a custom SLA. Model usage is included, or you bring your own key.
What if we've already started building?
Then compare what's left to build with what Kelu already does — usually connectors, sync, abstention, analytics and the Slack, Teams and helpdesk surfaces. You don't have to switch all at once: connect your sources to Kelu, then move one surface at a time, or call Kelu's API and MCP server from what you've built.
Can we bring our own model?
Yes, on every plan. Bring your own model key sends answers through your own OpenAI, Anthropic, Google Gemini or AWS Bedrock account, under your agreement with that provider. You keep the model choice and the invoice; Kelu is the platform.
Is the retrieval really different from a build on OpenAI file search?
File search is a solid base: it combines semantic and keyword search with reciprocal rank fusion and lets you set a ranker and a score threshold. The differences are around it. Kelu cleans pages into markdown and chunks them by heading rather than by token count, rewrites the question with an LLM before searching, reranks with a cross-encoder, and then decides whether the sources cover the question at all — and if not, answers that it doesn't know. It also keeps 24 source types in sync, which file search leaves to you.
See it on your own docs
Connect a source and get your first cited AI answer in under 5 minutes. Free plan, no credit card required.
Free to start · No credit card required · Talk to Sales →