Skip to content
ESEN
Contact

Technology · Integration

Generative AI integration for your company

A language model is useful when it works inside a specific task: answering with your documents in front of it, sorting what comes in or leaving a draft for someone to review. We integrate those features into your website and your tools, with sources in view and a person wherever a mistake is expensive.

Tell us about your project See an example

Often combined with

Fig. 01Archive and core12ANSWER · WITH SOURCESLooping
The query opens the drawers it needs, the core prepares the answer with its references, and the push button lets it through or sends it to review.

Situations

What you can solve by integrating generative AI

This isn't a chatbot that knows everything. It is a feature with one job, a set of sources and a clear limit on what it may decide alone.

  • Questions answered by searching documents

    Manuals, contracts, procedures, spec sheets. An assistant answers citing the document each answer comes from.

  • Messages and documents that need sorting

    Emails, tickets or forms someone classifies by hand. The model proposes the category and the key data; anything doubtful goes to review.

  • Text written over and over

    Replies, summaries, product descriptions. They are generated as a draft, in your tone and with your data, and a person decides what goes out.

  • An AI feature inside your product

    Search that understands the question or a field that fills itself in, built into the application your customers already use.

The review control: a person decides what goes out

What to check before starting

  • Which documents or data the model may read, and which never
  • Which mistake is acceptable and which isn't
  • Who reviews, and at which point
  • Which model provider your data policy allows

Services

What we build with generative AI

Specific features integrated where the work already happens. Each one is delivered with the means to check whether it answers well.

  • Module 1 · Archive

    Document lookup

    Assistants that search your documentation and answer with references, limited to the sources you decide.

    Examples of deliverables

    • Assistant over manuals
    • Contract search
    • Internal help for support
  • Module 2 · Core

    Classification and extraction

    Reading emails, forms and documents to assign a category, extract fields and leave empty what isn't there.

    Examples of deliverables

    • Ticket classification
    • Invoice data
    • Request routing
  • Module 3 · Answer screen

    Built-in drafts

    Generating replies, summaries and text inside the tool where the work is done, always as a proposal.

    Examples of deliverables

    • Draft reply
    • Case summary
    • Product description
  • Module 4 · Review control

    Review and control

    Review screens, a log of what was proposed and what was accepted, and limits on usage and cost.

    Examples of deliverables

    • Review queue
    • Answer log
    • Evaluation with real cases

A working example

A query answered from sample documents

A simulated demo. Pick a question: the answer shows which sample document each statement comes from. One of the questions has no answer in the documents, and the assistant says so.

Simulated demo · sample data

Pick a question

Assistant's answer

The answer will appear here, with the number of the document each sentence relies on.

Sample documents

  1. 1
    Terms of sale

    The customer has 30 calendar days from delivery to return the product, provided the original packaging is kept.

  2. 2
    Spec sheet · model K2

    The K2 model supports a maximum load of 120 kg and ships with two batteries.

  3. 3
    Warranty policy

    All models carry a 3-year warranty. Batteries, 1 year.

There is no model behind this demo: the answers were prepared in advance to show the behavior. A real system can get things wrong; that is why it shows its sources and what matters goes past a person.

Integration

How it fits with your systems

The model is one piece among others. What decides the outcome is which sources it receives, where the answer appears and who confirms it.

Where it comes from

  • Your documentsmanuals, contracts and spec sheets
  • Email, forms and ticketswhat comes in
  • Data prepared in Python clean, chunked text
Generative AIAI featuresearches, writes and cites

Where it goes

Dependencies worth checking

  • Which model provider is used, and its terms on data usage
  • Which documents are current and which contradict each other
  • What each query costs and how many are expected
  • How we will measure whether the answers are good

How we work and what we deliver

How we run an AI integration

  1. 01

    Initial review

    The task, a sample of real documents or messages, and the mistake that must not happen.

  2. 02

    Scope

    What the model reads, what it proposes, who confirms and what is logged.

  3. 03

    First release

    The feature connected to a real sample, with sources visible in every answer.

  4. 04

    Validation

    Evaluation with questions and cases of yours, including the ones that have no answer.

  5. 05

    Documentation

    Which sources it uses, how they are updated, how to switch models and how to read the log.

What you get when the job closes

  • The code and the model's instructions, in your name
  • A set of evaluation cases
  • The log of queries and answers
  • A guide to updating sources and switching models

Questions

Questions about Generative AI

Missing yours? Write it to info@veodar.com and we'll reply with a specific question, not a brochure.

Are the answers always correct?

No, and be wary of anyone who promises that. A model can be wrong or fail to find the data. That is why every answer shows its sources, the assistant says when it doesn't know, and anything with consequences goes past a person.

Is the model trained on my documents?

It doesn't need to be. The usual setup searches your documents at query time and hands the model only the relevant passages. The documents stay where they are and can be withdrawn whenever you like.

What data is sent to the model provider?

The passages needed for each query. We choose a provider and terms that don't use that data for training, or a model you host yourself if your policy requires it. It is written into the scope.

Am I locked in to one provider?

No. The task, the instructions and the test cases are yours and can be taken to another model. The evaluation cases exist precisely to check that the switch doesn't make results worse.

How is this different from an AI automation?

Here the point is adding one specific feature to a website or a tool. An automation is a complete flow over your systems, with its credentials and its confirmation step. Often the first is a piece of the second.

Updated:

Related services

All services

Other technologies

What it is combined with

All technologies
  • Python programmingImporting, cleaning and matching data, generating reports and automating repetitive tasks. Tools shaped around your files and systems.
  • Node.js backendAPIs, connections between systems and background processes. We automate the exchange of information between your website and your business tools.
  • Next.js developmentApplications that combine public pages, user sign-in and server functions. Development and integration of the different parts of the project.

Get started

Tell us which task you want to speed up.

What comes in, what should come out and which mistake must not happen. A couple of real examples tell us whether it makes sense.

What helps us assess it

  • Which task it is and who does it today
  • Real examples of what comes in
  • Which documents or data it would need
  • Which mistake must not happen
Tell us about your project info@veodar.com

Before you write, see how we price a project or our solutions by department.

Fig. 07Morse keyH · E · L · L · O