Agent and knowledge platform
Retrievix turns your documents into chat apps, agents and workflows that cite their sources. Publish each one as a web app, a widget, an API or an MCP server, all from your own private workspace.
How it works
Upload documents and choose how they are split into chunks. Retrievix embeds and indexes them for hybrid search, and you can test retrieval before any app depends on it.
upload, chunk, index
Start from a chat app, an agent or a blank canvas. Attach knowledge, tools and a model, then try it in a debug pane that streams as you type.
prompt, tools, model
Switch on a web app, a widget, an API or an MCP server. Every conversation is logged from the first message, so you can see what people asked and correct what it said.
share, embed, call
The product
Knowledge feeds apps, apps grow into workflows, and every one of them can be published. The drawings below use example names, not customer data.
01Knowledge
Upload files or sync a data source, then decide how each one is split. Every chunk can be read and edited, and a retrieval test shows exactly which passages a question would pull back.
02Apps and agents
A chat app answers from the knowledge you give it. An agent decides which tools to call and calls them in turn until it has an answer. Both stream in a debug pane while you tune the prompt.
03Workflows
Place nodes on a canvas and wire them together. Branch on a condition, loop over a list, run code or call an HTTP endpoint. A test run shows each node as it runs, with its inputs and outputs.
04Publishing
Each app can publish in four ways, and you choose which are on. Share a link to a hosted web app, drop a widget on your site, call it from your own code, or serve it to any client that speaks MCP.
Capabilities
Each of these is a setting or a screen in the product, not a promise on a roadmap.
General or parent child chunks, your own separators, sizes and overlap, with a preview before anything is indexed.
Vector and full text search together, with optional reranking and a score threshold per knowledge base.
Answers carry the documents and passages they drew on, so a reader can check the source.
Ask a knowledge base a question and see which chunks come back and how each one scored.
Add fields to documents and narrow retrieval on them, by hand or from the conversation.
Built in tools, custom OpenAPI tools, workflows published as tools and tools from MCP servers.
About twenty node types, including branches, loops, code and HTTP requests, with a live trace on every test run.
Variables in a prompt become inputs on the published app and fields in its API.
Every conversation and run is kept. Correct a reply with an annotation, and with annotation reply on, similar questions get the corrected answer.
Messages, users, tokens and cost for each app over the time range you pick.
DeepSeek is connected by default. Other providers, including OpenAI compatible endpoints, install as plugins.
Every account gets its own workspace, so its apps, knowledge, tools and logs stay its own.
Publishing
Publishing is a set of switches on the app, not a second project. The snippets below point at an example host, and yours will carry your own domain and codes.
Every app gets its own page on your deployment. Give it a title, an icon and a description, then send the link. Turn the page off and the link stops working.
https://retrievix.example.com/chat/k3JdQ9vLmP2xQuestions
Short answers about models, data and sharing.
DeepSeek is connected as the default chat model. Other providers install as plugins, including any endpoint that speaks the OpenAI API, and each app picks its own model.
Embedding and rerank models are configured the same way. A knowledge base needs an embedding model before it can index for vector search.
Each account works in its own private workspace. Its documents, indexes, conversations and logs are kept apart from every other account.
To answer a question, the question and the passages retrieved for it are sent to the model provider you configured. Choose a provider whose terms you are happy with.
Anyone can create an account. Each account gets a private workspace and a daily message allowance.
Anyone with the link can use a published web app or widget without an account.
Access to a published app can be switched off at any time, and its API keys can be revoked one by one.
The Model Context Protocol is an open standard that lets an assistant discover and call tools on a server. Any Retrievix app can be served as an MCP server, so a client that speaks the protocol can call it as a tool.
It works the other way too. Agents and workflows in Retrievix can use tools from MCP servers you add.
Yes. Each app has its own API keys and endpoints for chat, completion and workflow runs. Responses can stream as they are written, and every call appears in the app logs.
On one machine you run. The backend is a set of containers started together, and this web app sits in front of them on the same origin.
Create an account, add a knowledge base and test a question against it. Then build the app that answers it and publish it.