Vault RAG

Our Retrieval-Augmented Generation solution

A local RAG solutions that keeps the company documents totally secret, allowing researches into the documents without any access to the WEB. Like in a protected Vault.

VaultRAG, our AI Retrieval-Augmented Generation solution is an advanced document query platform based on local LLM models, and so able to analyze sensitive documents avoiding any access to the Internet and keeping the company documents totally secret. This is the key aspect to highlight from the start: VaultRAG runs entirely on-premises. This means your corporate data never leaves your infrastructure, ensuring full data privacy and control.


As the second key feature, we underline that the VaultRAG platform supports an unlimited number of documents and users, allowing you to build a scalable and unrestricted knowledge base.

Moreover, you do not need to pay any software license, fee, online subscription… The VaultRAG is an open-source solution built mainly using python code. It can be installed on hardware of any size, and more than one copy.

Once logged in, the chat interface is designed to be clean and intuitive. It's a modern single page interface powered by javascripts that manage the DOM structure of the HTML page to add and delete objects. In this way, the interaction with the back-end server is minimized, is fast and pleasing, and no other HTML page is opened during the interaction.

On the left of the interface, you’ll find the document repository, organized into several database/folders of documents, which are a collection of personal or corporate documents available to be analyzed.


Let's imagine a corporate with many thousands of documents to be analyzed. A RAG system of any kind needs to load them into a vector database to assist the future queries. Documents (usually PDF) need to be converted into text and indexed in a specific Vector database. It's a time-consuming activity, loading the documents from the original storage into Vector databases.
VaultRAG stores permanently the corporate documents into several Vector databases, or digital collection of documents. Corporate users will find them already available in the ValutRAG, organized into logical folders to reduce inquiry time avoiding to query too many thousands of documents at the same time. But also organized to match the corporate organization, their needs, and defining secure access system based on the organization.

The folders of documents are persistent, and administered by authorized personnel. Anyway, also the user can add new documents into personal VaultRAG folders. They are persistent, too. He/she does not need to load them every time. They are always available to be analyzed by the user who loaded them.

At the center of the VaultRAG is 'Ratio Expert', our AI-powered assistant. You need first to select at least a document folder with the collection of documents to query. Then you can ask what you like to the 'Ratio Expert' about the documents included in the folder.

Now the AI model analyzes the documentation and provides a clear, structured, and detailed answer. Tipically, it summarizes all the information found in the documents about your query in a short description of few dozens of text lines. What truly sets the system apart is its reliability: every statement is backed by a source document reference including the page number. By clicking on the references, VaultRAG takes you directly to the exact page of the source document where the information was extracted.

This ensures full traceability and verifiability.


Switching to the document management, by clicking on 'View/Filter Documents', you can access the document repository to select specific documents to be enquiry from the documents available in the folders. Using the search bar, you can search by loading date or period, pubblising date, text name, category and so on. This allows you to isolate the relevant documents for your research, alsomaking it easy to narrow down results to specific domains or topics. Once you identify the documents, you can select them in bulk and confirm to make them available to restrict querying with the Ratio Expert.

The available corporate documents in the repository and folders are loaded and organized by a system admin, or some experts authorized to do it. But you can load also additional personal documents to be analyzed and add some personal folders that will be available only to you even in the future.
To expand your knowledge base with a new folder storing a collection of your personal and reserved documents, you need to create a new folder just by clicking on 'Create New Database', and then name it what you prefer." Within this new folder, you can upload personal documents from your local machine.
VaultRAG offers two different document processing approaches depending on your needs: text based or with OCR capability. We can either extract text from PDF documents while excluding visual content, or enable advanced extraction with OCR capabilities that processes both text and embedded images, allowing the AI to capture richer contextual information. For example, if you select an organizational chart in PDF format, an advance OCR loading is better.

The system always processes and indexes the documents to be loaded within seconds, notifing you with a confirmation message that the upload was successful. In this way, as with common documents, you can ask what you like also about information written into images and schemas.


In addition, VaultRAG maintains a searchable chat history, allowing users to revisit previous questions, review answers, and quickly recover past insights.

Not less important advantage is flexibility. VaultRAG can be fully customized to match specific business needs, from the underlying AI model to the interface design, and even the language of the AI assistant itself.

It can use more than an LLM model, optionally also online models if requested for some specific less private documents.

It can be integrated with the corporate systems, database, applications, AI agents, and more.

With VaultRAG, your company’s knowledge is always accessible through a simple chat interface—secure, scalable, and fully verifiable.

Look at the demo shown in the video below to see how you can operate with our VaultRAG.
Note: The application is continuosly improved. The video shows only features developed several weeks ago.

Download English version of the video(about 151 MB)

All our recent videos are available also from our Youtube channel Rationence.

SQL ChatBot! Update 2025-09

A new powerful chat bot based on a local and private LLM model of Artificial Intelligence, able to translate into SQL your free data analysis prompt, and then applying them to any db to provide insight, SQL code, profile and export data!

The video below illustrates the main features of our SQL Chat Bot, friendly called "chattino".

This last and enhanced version provide you:

1. a new database selector among predefined database such as in SQLite, Oracle or SQL Server.
2. the possibility to select and load into your database data files in CSV format.
3. a column description and a table preview of the data loaded or available in a schema of your database.
4. features to extract and export information from your tables or to profile them with many charts including bar charts, word cloud, box-plot...
5. features to select and order the tables that you consider as input for your queries, this order is saved and restored per each available database.
6. a chat interface in any natural language, being able to translate user requests into queries and actions, automatically submitted and executed on the database.
7. a program in SQL corresponding to your request to be executed or submitted.
8. the program in SQL can be executed immediately or first controlled / modified / verified and then submitted.
9. produced queries can be slightly different depending on the SQL syntax specific of the selected database engine.
10. the results of the execution or submission of the code starting from your data and database are shown immediately together with the elapsed time of execution and possible messages.
11. an optional extraction of your data as requested to be exported into CSV or Excel files.
12. some optional charts can be requested to profile or describe your data.
13. when available by the current database, functions and statistics such as correlation can be provided too.
14. the SQL programs produced by our Chattino are always well indented and commented.
15. the SQL programs produced can be cumulated in a job, copied, modified and then stored as you like.
16. a second chat interface in any natural language can be provided as a premise for the LLM model, to add specific information about complex relations among tables that can improve the query production. Usually provided by your DBA.
17. the possibility like this demo to run the Chattino on a local and private LLM model of Artificial Intelligence, without any access to external provider of AI, without license, without internet access of any kind. So, with no risk to spread information outside your organization.
18. the Chattino interface are web based and can be shared among many users in your organization.

This ChatBot is written mainly in python and do not require any specific software license or access to AI providers.


Looking at the demo shown in the videos below, you can observe that this new version of our SQL Chattino is very clever in interpreting your requests starting from prompts of few words, even just a line. It can produce complex SQL query, with any possible clauses, even with nested subqueries and aggregations. It is able to join or merge multiple tables without any user's hints or external metadata. It can also provide DDL or DML code to create tables, insert data, modify date, delete date etc... And often it can correct your query request, and find the right table and variables to be used by interpreting your prompt in an unexpected and clever way.
Finally, our Chattino SQL automatically detects all the keys that link the available tables if the variables have the same name, and sometimes detects more complex association among data by itself, even if the variables have different names.

Here below you can find a short video of our "Chattino SQL". About 19-20 minutes and not enough to illustrate all its features.

Download English version of the video(about 192 MB)

Here below you can find a long video of our "Chattino SQL" in Italian. About 40 minutes and not enough to illustrate all its features.

Download Italian version of the video(about 312 MB)

All our recent videos are available also from our Youtube channel Rationence.

Python and SQL ChatBot!

More and more in demand, chat bots are now able to translate your data analysis prompt into Python and SQL code, and then applying them to provide insight, code, export data, and draw charts!

Here below, we are showing some custom chatbot solutions that demonstrate our ability to develop web apps that starting from data analysis requests in a natural language of your choice are able to interpret your prompt and execute programs on your data without coding any statement.
They are under continuous development to improve their funcionalities. The versions here illustrated are currently available in Italian and English languages. Demo and screenshots were built at the state of the art in date 14th April 2025.

These solutions can provide you:

1. a column description and a table preview of the data loaded or available in a schema of your database.
2. a chat interface in any natural language, being able to translate user requests into queries and actions, automaticaly submitted and executed on the database.
3. a program in SQL or Python code corresponding to your request to be executed or submitted.
4. an explanation of the steps to produce the code above.
5. the results of the execution or submission of the code starting from your data and database.
6. an optional extraction of your data as requested to be exported into CSV or Excel files.
7. some optional charts as requested to describe your data.
8. functions and statistics such as correlation can be provided too.

These ChatBots are written in python and do not require any specific software license.
By default, an API KEY is required to access an LLM model such as those provided by Open AI.
However, alternative open-source LLM Models can be implemented on a local private server to avoid any costs, and to prevent any even small amounts of information from being collected by providers such as Open AI.

Looking at the demos shown in the videos below, you can observe that these ChatBots are quite intelligent in interpreting your requests starting from prompts of few words, even just a line.
Finally, they automatically detect all the keys that link the available tables, and they are able to join or merge multiple tables without any user's hints or external metadata, as to code subqueries by themselves.

Below you can find a short video of our chatbot demo named "Chattino", version Python coding.

Download English version of the video(about 32 MB)

Download Italian version of the video(about 39 MB)

Below you can find a short video of our chatbot demo named "Chattino", version SQL coding.

Download English version of the video(about 17 MB)

Download Italian version of the video(about 31 MB)

CV Ranking with Large Language Models

A smart solution to support the HR function in scouting the right candidates among a large set of CVs

Currently, personnel selection processes require significant amounts of time and effort, especially for companies managing multiple open positions simultaneously or receiving large quantities of CVs after posting a job offer.
CV-Ranking is a smart solution based on LLM models to support the HR function in scouting CVs, in order to find the most suitable ones for the required position. The system takes as input a natural language job description and a large set of resumes as input, and then provides a ranking of the most suitable CVs for the entered job description, speeding up the CV selection process.

CV acquisition module takes PDF formatted CVs as input and extracts the textual part, making the documents processable. Then a pre-processing phase follows, which aims to clean and process the textual data previously extracted from the CVs, through Lemmatization, N-grams and Bi-grams creations, stop words removal and so on.


Based on the results of the aforementioned phases, a word embedding process is carried out, which is a technique that allows storing both semantic and syntactic information of words starting from an unannotated corpus and building a vector space where word vectors are closer if the words occur in the same linguistic contexts, meaning they are recognized as semantically similar. The vectors resulting from this process are inserted into a knowledge base, preserving the metadata, in order to subsequently trace back from the vectors to the CVs that originated them.


Then follows a Job Description analysis module, whose purpose is to extract skills from the Job Description entered by the user in any natural language to be searched among the CVs. The extraction is performed by providing the relevant text to a Large Language Model (LLM) accompanied by an appropriate prompt.

The vectors generated by the word embedding of CVs and the skills extracted from the job description are then crossed, evaluating their similarity, under the assumption that, due to the functioning of the applied word embedding procedure, numerically similar vectors represent texts similar in meaning.

At this point, for each document in the knowledge base are associated N similarity scores, where N is the number of skills of interest. These N scores are then summarized into a single score through a weighted sum, which serves as a model to provide an overall ranking related to all skills, constituting the system's output.

The current CV-Ranking interface allows the user to customize some of the previously illustrated phases according to their needs. By the way, Large Language Models might have "hallucinations", meaning it might provide inconsistent outputs. The remedy devised for this intrinsic and unpredictable possibility is to provide a feature that allows users to select and remove any undesired skills via a dropdown menu based on their needs.

Not all skills extracted from a Job Description might hold equal importance for a user. Therefore, a functionality was incorporated into the interface to enable the selection of skills deemed strictly necessary for the desired candidate.

Once these specifics are entered, it is then possible to trigger the analysis by pressing a button and obtain the result, consisting of a table in descending order showing the top ten CVs most suitable for the entered Job Description, along with their respective scores and the option to view the CVs of individual candidates.
Simultaneously with the tabular output, the user is presented with graphs that allow visualization and comparison of the details of the scores generated on the top CVs for each individual skill used in the search.

The system has been tested with more than 2500 CVs, providing results in just a minute.