🧠 OpenRAG
OpenRAG is a RAG platform that cleans and structures a business's content, indexes it on our own server, and makes it queryable in natural language by external AI systems through a single permissioned door.
Vectors are not published; ACCESS is opened.
This is a product introduction, not a list of promises. Below, the parts that work today and the parts still on the roadmap are marked separately. No dates are promised for roadmap items; as they are completed they move into the "working today" section.
The core principle
OpenRAG's entire design rests on a single sentence: vectors are not published, access is opened. What is given to the outside is not a copy of the content, but the right to ask questions of it.
A vector is an engine, not a product
A vector is specific to the model that produced it; another AI cannot consume it as-is. It also inflates the text many times over. That is why vectors are never handed out.
What is opened is query access
A question comes in; an answer and the source passages it rests on go back out. The full content is never copied, distributed or downloaded.
Permission belongs to the business
The business decides which content becomes externally queryable; participation is opt-in. The permission and consent layer is on the roadmap.
How it works — target architecture
The cleaning, structuring, indexing and search links of this flow work today. Self-hosted embedding with reranking, and the single door to the outside world, are still on the roadmap; each step below states which is which.
Sources are collected
xloji sites, uploaded documents and crawled web pages are the content sources. (A WordPress plugin and a public ingestion API are on the roadmap.)
It is cleaned and structured
Messy content is cleaned, split into meaningful chunks and brought into a form RAG can use. This step works today.
It is indexed and searched
Chunks are turned into vectors and written to pgvector; they are found via an HNSW index and hybrid search (keyword + vector). This step works today; adding the reranking link is on the roadmap.
It is reached through a single door
External AI systems ask in natural language through a single permissioned door; answers and source passages come back, not the vectors themselves. This door does not exist yet; it is on the roadmap.
Parts that work todayLive
The following are running on the platform today and are already used by the AI tools on xloji.com.
pgvector + HNSW index
Content chunks are stored as vectors in the database and searched with an HNSW index over cosine similarity.
Hybrid search
Keyword search and vector search run together; the two result lists are fused into a single ranking. This finds both exact term matches and semantically close passages.
Multi-tenant isolation
Every business, site and author has its own RAG. A query runs only against its own tenant's index; data never mixes across tenants.
Document cleaning and RAG structuring
Uploaded documents are cleaned, chunked and structured for RAG. The Data Structuring tool on the platform is the user-facing face of this pipeline.
Web crawling
A site's pages are crawled, their content is extracted and passed through the same cleaning and indexing pipeline.
Public RAG feed in custom AI tools
A custom AI tool you create can be fed from a public RAG source; the tool grounds its answers in that knowledge base.
File and media storage
Files and media live in object storage — not in the vector store. The RAG record points back to the original file.
Embedding and reranking on our own servers
Embedding and reranking now run on our own servers in a separate service. Text is turned into vectors without leaving the server, and candidate results pass through a local reranker to improve accuracy.
Query rewriting
A vague or incomplete question is tidied up before the search runs. Users do not have to phrase the "right" question; the system makes the query search-ready.
Text from images, audio and video
Text extraction from images and transcription of audio and video run with local tools on the server, and the resulting text enters the same cleaning and indexing pipeline. Data does not leave the server.
What is on the roadmapPlanned
The following do not exist yet. They are planned and designed items; this page does not present them as if they were already built.
One door: MCP / API gateway
A single gateway that external AI systems can query in natural language. The endpoints are written and tested; the door is kept CLOSED today and opened only for supervised tests. Public opening follows the consent layer.
Permission and consent layer
A business joins by its own choice. Key issuing and site connection already work; what remains is rate limiting, abuse protection, and letting a business manage its permissions from a single screen.
AI visibility layer
llms.txt and schema.org markup are live — on the main site and on sub-sites. What remains is the /.well-known discovery files and publishing MCP registry information.
WordPress plugin
The plugin is written and runs on a real WordPress install: the site connects to an xloji.com account, content is sent one-way, and the key never passes through the user's hands. The remaining step is submission to the wordpress.org repository.
We do not give dates for these items. When an item is finished it moves up into the "parts that work today" list, and this page is updated accordingly.
🔒 Data does not leave the server
OpenRAG's goal is to run every step of the RAG pipeline on our own server with local, free tools: embedding, reranking, text extraction from images and audio transcription. No media entering RAG is sent to an external service — this is a non-negotiable design rule.
The honest status: some of these steps are still performed by external services today. Moving them in-house is the first item on the roadmap.
How growth is planned
Vector search gets more expensive as content grows. OpenRAG's design limits that cost from the start in three ways.
There is no single giant index
Each business has its own small index and a query runs only against that index. Even as total content grows, the area a single query scans stays small; search cost scales with that one business's size, not with the platform's total.
Vector size and precision are adjustable
Using shorter vector dimensions and more compact number formats, the same content takes up markedly less space. These choices are built into the design from the start so that no expensive migration is needed later.
Cold tier
The index of a business that has not been queried for a long time is moved to cheap persistent storage and pulled back when a query arrives for it. That way only currently active content stays in fast memory. This tier is on the roadmap.
Frequently Asked Questions
What exactly is OpenRAG?
It is a RAG platform that cleans and structures a business's content, indexes it on our own server, and makes it queryable in natural language by external AI systems through a single permissioned door. It is not a chat tool; it is the knowledge layer that runs underneath such tools.
Do you hand my vectors or my content to anyone else?
No. The core principle is: vectors are not published, access is opened. What goes outside is not a copy of the content but the right to ask it questions; only the supporting passages come back with the answer. On top of that, the business decides which content becomes externally queryable.
What can actually be used today?
The vector database and index, hybrid search, multi-tenant isolation, document cleaning and RAG structuring, web crawling, the public RAG feed in custom AI tools, and file/media storage all work today. Self-hosted embedding, reranking, query rewriting, the external door, the permission layer, the visibility layer and the WordPress plugin do not exist yet.
Does my data leave the server?
The goal is for every step of the RAG pipeline to run with local tools on the server, with no media entering RAG ever sent to an external service. The honest status: some of these steps are still handled by external services today, and moving them in-house is the first item on the roadmap.
When will the roadmap features be ready?
We do not promise dates. That is exactly the purpose of this page: to show separately what works today and what is planned. When an item is completed it moves into the "parts that work today" list.
Is my content used or sold for AI training?
No. There is no such use today. The idea of a permissioned text corpus is the furthest-out, legally heaviest item on the roadmap; even if it happens, no content would be used for it without the business's explicit consent.
Head back to the AI portal for the other AI tools.
Frequently Asked Questions
What is OpenRAG?
OpenRAG is a RAG (Retrieval-Augmented Generation) system and a single-door AI gateway running on xloji.com's own server. When a question arrives, it first finds the relevant content, then generates an answer based on this content. It also serves as a central gateway allowing AI tools outside of xloji.com to access xloji.com's information from a single point.
What does RAG mean and how does OpenRAG implement it?
RAG stands for 'Retrieval-Augmented Generation', meaning it searches for information from relevant sources before an AI generates an answer, then creates a response based on that information. OpenRAG runs these two steps, the search and generation steps, together on its own server. Thus, the answers given by OpenRAG are not random, but based on real content belonging to xloji.com.
What does 'single-door AI gateway' mean?
A single-door AI gateway means that different AI requests and integrations are passed through a single central entry point instead of separate systems. OpenRAG takes on this role: both xloji.com's own 'Ask a Question' feature and AI tools connected from outside receive service through the same OpenRAG infrastructure. This central structure allows AI-related traffic to be managed from a single point.
What difference does it make that OpenRAG runs on its own server?
OpenRAG runs on xloji.com's own server without relying on a third-party cloud service. This means that the searched and processed data remains under xloji.com's own control. As a structure running on its own server, OpenRAG performs search and answer generation operations within its own infrastructure without transferring data to external systems.
Who uses OpenRAG?
OpenRAG is used by three different groups: end-users who visit xloji.com and type into the 'Ask a Question' box, systems that query xloji.com's own RAG infrastructure, and external AI tools that access information about xloji.com via OpenRAG. For the end-user, OpenRAG returns an answer based on xloji.com content to the question asked. For external tools, OpenRAG is a context-free gateway used to retrieve information from xloji.com.
How does the 'Ask a Question' box on xloji.com work with OpenRAG?
A question entered into the 'Ask a Question' box on xloji.com is processed by OpenRAG. OpenRAG first searches for content related to the question within xloji.com's own data, then generates an answer based on the content it finds. This ensures that the answer given to the user is supported by real content from xloji.com.
How do external AI tools obtain information via OpenRAG?
OpenRAG serves as a single-gate passage enabling AI tools outside of xloji.com to access information about xloji.com. These tools can connect via OpenRAG and receive self-contained, understandable answers from xloji.com's content. Therefore, the information provided via OpenRAG is designed to be meaningful on its own, without requiring a separate context.
What are the limitations of OpenRAG?
The answers provided by OpenRAG are limited to the content provided to it; OpenRAG cannot produce a definitive answer about information that is not in the system or has not been added. OpenRAG is designed to remain limited to the content it has, rather than fabricating non-existent information. Therefore, it should be remembered that an answer obtained from OpenRAG is limited to the content that xloji.com currently possesses.
Is OpenRAG a chatbot?
OpenRAG is not a standalone chatbot in the classical sense; it is the RAG and gateway infrastructure that powers xloji.com's 'Ask a Question' feature and external tools' access to information. OpenRAG's core function is to find content corresponding to a question and generate an answer based on that content. As such, OpenRAG operates more as the infrastructure layer behind the answers than as an interface that directly chats with the user.
What is the purpose of OpenRAG for xloji.com?
The purpose of OpenRAG is to enable both end-users and external AI tools to access information belonging to xloji.com through a reliable and self-controlled infrastructure. Running on its own server ensures that this access happens without dependence on third-party systems. Thanks to its single gateway structure, OpenRAG combines different access points in a single central system.
Detailed information about OpenRAG
OpenRAG is a RAG (Retrieval-Augmented Generation) system running on xloji.com's own server, and it is also an AI gateway that passes AI-related requests from a single point. The 'Open' in its name indicates that the system offers an open and accessible structure, while 'RAG' refers to its method of operation, which involves first finding relevant data and then generating an answer based on that data. A question typed into the 'Ask a Question' box on xloji.com is first searched for relevant content by OpenRAG, and then an answer is generated based on this content. The same infrastructure also allows AI tools outside of xloji.com to access information belonging to xloji.com through OpenRAG and use this information in their own answers. This structure enables xloji.com to provide services through a single central infrastructure instead of setting up separate AI solutions for its different products.
The basic logic of the RAG approach is that an AI system, instead of relying solely on the general knowledge it was previously trained on, first retrieves up-to-date and accurate content from a source related to the question and then answers in light of this content. OpenRAG operates by combining these two steps – retrieval and generation – on its own server. This approach aims to reduce the tendency of AI models to sometimes generate non-existent information; because the answer is derived from actually found content, not from the model's own memory. In this way, the answers given by OpenRAG are based on xloji.com's own content, not on random or outdated information. Thus, the source of an answer produced by OpenRAG remains traceable as xloji.com's own data. The server belonging to xloji.com means that the searched and processed data is kept under xloji.com's own control without depending on a third-party cloud service.
The 'single-door AI gateway' side of OpenRAG means that different AI requests and integrations are routed through a single central entry point instead of separate systems. In this way, both xloji.com's own 'Ask a Question' feature and external tools connected via OpenRAG use the same infrastructure and the same basic logic. As a central gateway structure, OpenRAG ensures that AI-related traffic passes through a single point; this increases both the consistency of answers and control over the data as a system running on its own server. This single-point approach also allows maintenance and updates to be done in one place instead of in distributed systems. In this respect, OpenRAG is not only a question-and-answer tool, but also an infrastructure layer that unifies all of xloji.com's AI access points.
OpenRAG is used directly or indirectly by three different user groups: end-users who visit xloji.com and type into the 'Ask a Question' box, systems that query xloji.com's own RAG infrastructure, and external AI tools that access information belonging to xloji.com via OpenRAG. All three user groups receive answers fed from the same OpenRAG infrastructure, the same content pool; therefore, a consistent source of information is reached no matter where questions about xloji.com are asked. The answers given by OpenRAG are limited to the content provided to it; that is, a definitive answer should not be expected from OpenRAG about information that is not in the system or has not been added to its server. This limitation also forms the basis of OpenRAG's reliability: instead of fabricating non-existent information, it aims to produce answers limited to and based on the content it has.