Retrieval-augmented

Answers grounded in your own content

Every reply is built from your indexed documents and carries numbered citations back to the source. No context, no answer — the assistant says "I don't know" instead of inventing one.

Try the assistant Read the docs

The 60-second pitch

How it works, end to end

From a visitor's question to a cited answer — and onto any website in one line.

Myra Technolabs presents

Contexa

An embeddable assistant that answers only from your own content.

01 / 07

The product

Contexa, by Myra Technolabs — an embeddable assistant that answers from your own content.

0:00 / 0:58

~18 KB

Widget bundle, gzipped

1 line

To embed on any site

100%

Answers carry citations

Hallucinated sources

Features

Everything a client site needs, nothing it doesn't

Built as a focused product rather than a demo: theming, citations, multi-tenancy and reconnection are all handled for you.

Grounded, cited answers

Replies are assembled only from retrieved chunks. Each claim carries a [n] marker that links straight back to the source document.

Says "I don't know"

When retrieval comes back empty the widget short-circuits before the model is called, so it never invents an answer from nothing.

White-label theming

One hex value recolours the launcher, bubbles, gradients and focus rings. Driven by CSS variables a host page can override.

Realtime over Socket.io

A typing indicator appears the instant the server accepts the question, and the socket reconnects on its own if the link drops.

One line to embed

A single script tag with the site token. The chat runs in an isolated iframe, so the host page’s CSS and JS never collide with it.

Tenant-scoped retrieval

Each site token maps to one tenant on the server, and every vector search is filtered by it, so one deployment serves many client sites safely.

Score-thresholded recall

Weak matches are filtered before they reach the context window, which keeps prompts small and answers on-topic.

Bring your own stack

Qdrant for hybrid search, AWS Bedrock for embeddings and answers, MongoDB for transcripts. Models change in configuration, not code.

How it works

Live in three steps

Crawl your site

Start a crawl of the client website. Its pages are chunked, embedded and stored automatically, with source URLs for citations.

Create the site

Drop in the widget

Paste the script tag on the client site. The assistant is live, limited to that domain, with rate limits and daily quotas built in.

Read the integration docs

An assistant that cites its sources is one your support team can actually trust. Everything here is built around that single constraint.

Myra Technolabs — Intelligent Products. Automated Workflows. Real Business Impact.

Try it right now — bottom right

The launcher on this page is the same widget your clients would embed. Open it, ask something, and watch the citations come back.

Integration guide