Comparison

Kelu vs OpenAI Responses API

Last reviewed:

OpenAI shut down the Assistants API on August 26, 2026. Its replacement is the Responses API, with the Conversations API for threads and a hosted file search tool over vector stores. It is a set of building blocks for your own assistant. Kelu is a packaged knowledge assistant: connectors, retrieval, the widget, bots and analytics are shipped, not built. This comparison is about product vs building blocks, and what each costs to run.

Short answer

Choose the Responses API if the assistant is your product, or you need custom tools and flows, and you have engineers to build and run ingest, sync, UI, bots and analytics. Choose Kelu if you want a docs and support assistant that works now: 24 source types kept in sync, cited answers that abstain when your sources don't cover a question, and a widget, Slack and Teams bots, MCP and helpdesk surfaces included.

Feature comparison

Comparison based on OpenAI's own API docs, deprecation notices and pricing page. Time estimates for the build side are rough.

Feature
Kelu
OpenAI Responses API
Product type
Packaged knowledge assistant — works out of the box
API for building assistants and agents — you assemble the product
Time to first working assistant
Under an hour — connect a source, embed the widget
Days for a file search prototype; months for ingest, sync, UI, bots and analytics (rough estimate)
Source connectors
24 built in: websites, GitHub (files, code, issues, PRs, discussions), Notion, Confluence, Google Drive, OneDrive, Zendesk Help Center, Zendesk tickets, Jira, Jira Service Management, Linear, Salesforce Knowledge articles, Salesforce Cases, Discourse, Stack Overflow for Teams, YouTube, OpenAPI, S3, file uploads, curated Q&A
None for ingest — you crawl, convert and upload files to vector stores and keep them in sync. The model can call remote MCP servers for live data
Retrieval
Hybrid: pgvector + full-text, RRF fusion, LLM query rewrite, heading-aware chunks, cross-encoder reranking, coverage verdict
File search: semantic + keyword search merged with RRF, a ranker and a score threshold, over fixed-size chunks (800 tokens by default)
Source citations
Yes — numbered, renumbered to match the answer
File citations in the response; showing them is your UI's job
Abstention (refuse when sources don't cover it)
Yes — enforced: no answer and no citations
DIY — a score threshold drops weak chunks; deciding not to answer is your prompt's or code's job
Web widget
Yes — hosted, themeable, reCAPTCHA, white-label on Enterprise
ChatKit gives you an embeddable chat UI that you run with your own backend
Slack / Teams bots
Slack (shared-app install) and Microsoft Teams, on Pro and Enterprise
DIY — write the OAuth flow and event handlers for each platform
MCP server
Yes — per-knowledge-base endpoint
The model can call MCP servers; publishing your docs as one is yours to build
Support Ticket AI inside helpdesks
Yes — Zendesk, Intercom, Freshdesk, Front and Linear
DIY — write and maintain each helpdesk integration
Analytics
Yes — usage, feedback, citation coverage, coverage gaps and traces, on every plan
DIY — build the dashboard and gap reports
Total cost
Free / Pro $99 a month ($79 billed yearly) / Enterprise custom
Model tokens + file search ($2.50 per 1,000 calls, $0.10/GB/day storage after 1 GB free) + your engineering, hosting and upkeep

Where Kelu focuses differently

A product, not building blocks

The Responses API gives you a model, hosted file search and a chat UI kit. Turning them into a production assistant means building the ingest pipeline, the sync, the widget backend, the Slack and Teams bots, the analytics and the ops. Kelu ships all of that. The build-vs-buy question isn't 'can we build this?' — it's 'is this worth months of engineering, and an owner for years?'

Retrieval tuned for docs, out of the box

File search now does hybrid search with a ranker, which is a good base. What it doesn't do is prepare your docs: HTML→markdown that strips nav and boilerplate, code-block handling, chunks that follow headings rather than token counts, PDF structure, and a coverage check that decides whether to answer at all. Those are the pieces most build-your-own projects find they need after launch.

The connectors are the moat

Building one connector (say, Zendesk) with correct pagination, incremental sync and removal handling is weeks of engineering — a rough estimate. Doing 24 of them, and keeping each working as vendor APIs change, is a standing job. Kelu ships and maintains them, skips unchanged pages, and holds back large removals for review.

Keep your model and your bill

Bring your own model key (on every plan) means Kelu calls OpenAI, Anthropic, Google Gemini or AWS Bedrock on your account. You keep control over the model and the invoice; you skip owning the platform.

Got questions?

Frequently asked questions

Is the OpenAI Assistants API still available?

No. OpenAI announced its deprecation in August 2025 and removed it on August 26, 2026. The recommended replacement is the Responses API with the Conversations API; file search and vector stores carry over. Kelu was never built on the Assistants API, so the shutdown didn't affect it.

When would I still build on the Responses API instead?

If your assistant needs custom logic that isn't docs Q&A — a coding agent with your own tools, a multi-step workflow, an on-brand consumer character — the flexibility of building blocks is the right choice. For docs and support Q&A, a packaged product usually wins on time to value and ongoing upkeep.

Do I lose flexibility by using Kelu?

For custom agent flows, yes. For docs and support Q&A, much less — the hosted widget, a public REST and WebSocket API, and an MCP server give you enough surface to build your own experiences on top. React and JavaScript SDKs are on npm (@keluhq/sdk, @keluhq/react).

Can I bring my own OpenAI key?

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.

How does total cost compare?

Kelu is Free, Pro at $99 a month ($79 billed yearly), or custom Enterprise. Building on the Responses API means model and file search costs (with Kelu on your own key you pay the model costs too) plus engineering time — as a rough estimate, months for a production docs assistant with a widget and bots — plus ongoing upkeep as vendor APIs and models change.

See for yourself

Connect your docs and get your first AI-powered answer in under 5 minutes. No credit card required.

Free to start · No credit card required · Talk to Sales →