cloudparts

Kapa

A managed retrieval-augmented generation service for technical companies: it keeps an index of your docs, tickets, forums and code in sync and answers questions from it with citations, as an “Ask AI” widget, Slack or Discord bot, support-form deflector, API or MCP server.

What it is
Retrieval-augmented generation (RAG)
Used for
Documentation sites,Customer support
Type
Commercial product
Made by
kapa.ai Inc.
License
Proprietary
Facts checked
13 October 2026
Website

What it is

You connect sources, such as a docs site, GitHub, Zendesk tickets, Slack channels, Confluence, PDFs or an OpenAPI spec, and Kapa builds one search index from them. The ingestion layer does most of the work: it syncs only what changed, because re-embedding text, converting PDFs and annotating images for every refresh would be too expensive. New and edited content is picked up within minutes for most sources, deletions in a slower pass, Slack’s last seven days are re-scanned because its API cannot report changes, and websites are re-crawled daily, with any crawl that changes or drops a large share of pages held for human review instead of applied.

At question time Kapa retrieves the most relevant chunks and its prebuilt agents write an answer grounded in them, with citations, or say they don’t know. Those agents are deployed with one script tag (the website widget) or as Slack, Discord and Zendesk bots and a support-form deflector, and the questions people ask feed reports on coverage gaps in the docs. Kapa is model-agnostic and uses providers including OpenAI, Anthropic, Cohere and Voyage, plus some models of its own.

Developers can skip the prebuilt agents and use only the retrieval: an HTTP endpoint returns ranked chunks with no generation, in a fast default mode or a deep mode that runs a retrieval agent, and a hosted MCP server at <subdomain>.mcp.kapa.ai exposes the same search to Cursor, Claude Code or your own agent. A 14-day trial covers 1 index, 5,000 crawled pages, 5,000 retrieval queries and 1,000 answers; production use is a usage-based Growth plan whose price is quoted on request.

Why teams pick it

  • No RAG pipeline to run: crawling, PDF and image processing, chunking, incremental sync and evals are handled for you
  • Many surfaces from one index: website widget, Slack, Discord, Zendesk, support form, HTTP API and MCP server
  • Answers cite their sources, and question analytics show where the documentation is missing or unclear
  • Retrieval-only use is supported, so you can feed your own agent Kapa’s chunks instead of its answers

Watch out for

  • Closed and hosted: your content and users’ questions are processed by Kapa and its model providers (with DPAs and training opt-outs), and EU hosting and SSO are Enterprise features
  • Growth pricing is not published, and the free trial ends after 14 days
  • Default API limits are 60 retrieval requests and 100 document requests a minute per team
  • Answer quality depends on your docs: Kapa can only cite what its sources say, and stale or missing pages become “I don’t know” or coverage-gap reports

Alternatives

What all retrieval-augmented generation (rag) have in common →Compare side by side →
ElicitCommercialAn AI research assistant over 138 million academic papers and 545,000 clinical trials that searches, screens and extracts data from studies and writes reports with sentence-level citations, built for literature and systematic reviews.

How it fits with other technologies

Retrieve cited chunks from a Kapa project, with no generation · bash

curl -s -X POST "https://api.kapa.ai/query/v1/projects/$KAPA_PROJECT_ID/retrieval/" \
  -H "X-API-KEY: $KAPA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "How do I rotate an API key?", "mode": "deep"}'

# Returns the most relevant chunks, each with its source_url and content.
# The same search is a tool on the project's hosted MCP server: https://<subdomain>.mcp.kapa.ai

See it in a real stack

No architecture uses Kapa yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2019–2026

Emil Soerensen and Finn Bauer met in London and did master’s degrees in finance and computer science together; at Imperial College in 2019 they were already experimenting with early embeddings for question answering. In January 2023 two friends running tech companies independently told them the same thing: their support teams were swamped by technical questions their documentation already answered. Early deployments included a Discord bot for LangChain’s community of 30,000 developers.

They built a retrieval-augmented generation stack whose core, they wrote, was crawling infrastructure that keeps the source data current, and launched kapa.ai through Y Combinator’s Summer 2023 batch, with OpenAI, Airbyte and Next.js among early users. In October 2024 they announced a $3.2 million seed round led by Initialized Capital, with angels including the RAG paper co-author Douwe Kiela, Solomon Hykes and Amjad Masad, and more than 100 customers.

By 2026 Kapa was offering its retrieval directly to developers building their own agents, through an API, a hosted MCP server and LangChain and n8n integrations. In October 2026 it released Kapa MCP, which lets Claude, ChatGPT or Cursor set up and run Kapa itself, and published Company Knowledge Bench, a 1,000-question retrieval benchmark built from production data.

Why it’s built this way

Sync only what changed
Embedding, PDF conversion and image annotation cost money per item, so re-indexing whole sources would make even weekly refreshes too expensive; per-source change tracking keeps answers minutes old instead.
Gate website crawls before applying them
A redesign or a moved docs tree can silently break a crawl, so a crawl that changes or removes a large share of pages waits for human review rather than corrupting the live index.
Model-agnostic, evaluated in-house
The founders saw their success hinge on measuring answer quality, so they test models from several providers against their own evaluations and switch to whichever answers technical questions best.
Answer only from sources, with citations
Kapa optimizes for accuracy on technical questions, so answers are grounded in the connected sources, and when those don’t cover a question it says “I don’t know” rather than guessing.

Timeline

  1. 2019

    Founders experiment with embeddings for question answering at Imperial College London

  2. 2023

    Two founder friends ask for help with repetitive support questions in January

  3. 2023

    Launches through Y Combinator’s Summer 2023 batch

  4. 2024

    $3.2 million seed round led by Initialized Capital (October)

  5. 2026

    Kapa MCP and Company Knowledge Bench released (October)

Sources11 · facts checked 13 October 2026

Improve this page on GitHub