Softellar

RAG Development

RAG Development Services That Show Their Sources

We build retrieval-augmented generation into your products and internal tools. Every answer comes from your own documents, points to the passage it used and keeps each person’s access rights.

Discuss your documents

We reply within 24 hours

One answer, with its sources

Question

Can contractors get access to the staging environment?

Answer

Yes, with an access request signed by their hiring manager1. Their accounts are disabled on the last day of the contract2.

  1. 1Access policy

    Section 4.2, Contractors

    Contractor access to staging requires an access request signed by the hiring manager.

  2. 2Contractor offboarding

    Step 3, Accounts

    All contractor accounts are disabled on the last day of the contract.

When the sources stop

Do contractors get parking? The documents you may open do not cover this, so there is no answer to give.

Clutch rating, 12 reviews
5.0
Projects since 2021
70+

Companies we have built software for

  • Multi Radiance Medical
  • Loma Lux Laboratories
  • WellVu
  • Commercial Capital Group
  • Grand Auto
  • HostMy
  • Land Base
  • Auto Tuning

From a Question to a Sourced Answer

Documents go into the index on one track, and questions come out of it on the other. Each number is a step we build and your team helps decide.

Prepare your documents

  1. CollectSources synced on a schedule
  2. ReadScans and tables turned into text
  3. Split and indexChunks with their access groups
The indexPassages, their vectors, their sources and who may open them

Answer a question

  1. RetrievePassages the user may open
  2. DraftAn answer from those passages only
  3. Cite and checkSource marks, or a plain decline
  1. Prepare your documents

    Collect

    What we build

    • Connectors to the file stores, wikis and databases that hold the answers
    • Scheduled or event-driven syncs, so changed and deleted documents reach the index
    Person

    Which sources count as the truth, and how fresh each one has to be.

  2. Prepare your documents

    Read

    What we build

    • OCR and layout extraction for scans, so tables keep their rows and columns
    • Clean text with the title, headings and page numbers it came from
    Person

    Which document types matter, and what a well-read page looks like.

  3. Prepare your documents

    Split and index

    What we build

    • Chunks sized to your documents, each with its title, section and date
    • Access groups stored on every chunk, so permissions live in the index
    Person

    Who may see which source, and which fields stay out of the index.

  4. Answer a question

    Retrieve

    What we build

    • Keyword and vector search in one query, filtered by the user’s access first
    • A reranker that keeps only the passages worth reading
    Person

    Which questions matter most, and what a good passage for each looks like.

  5. Answer a question

    Draft

    What we build

    • A prompt that holds the question, the passages and their source IDs, nothing else
    • A private model endpoint, with token limits and a schema for the answer
    Person

    The tone, the length and when an answer must hand over to a person.

  6. Answer a question

    Cite and check

    What we build

    • Numbered source marks that open the exact passage the answer used
    • A plain decline when no passage covers the question
    Person

    Who reviews the questions the documents could not answer.

Where These Controls Run Today

The three AI products we built, then the analytics and data sync work behind the same access and freshness controls.

United States

Invoice recognition SaaS on Azure AI

Outcomes

  • Invoice processing up to 80% faster
  • Over 60% fewer manual correction requests
  • Accurate extraction across 50+ invoice templates
  • .NET
  • React
  • Azure Cognitive Services
  • Azure Service Bus

Inside the Retrieval Step

Most wrong answers start with the wrong passages. Here is what happens between the question and the prompt.

A user asks a question
  1. Access filterOnly chunks the user may open
  2. In parallel

    Keyword searchExact terms, codes and names
    Vector searchThe same meaning in other words
  3. Fused rankingOne list from both, by RRF
  4. RerankerTop results rescored by meaning
  5. PromptTop passages with their source IDs
How these steps workFilters in vector queriesHybrid search in Azure AI SearchSemantic ranking in Azure AI Search

Clients on Their Data Work

A unified analytics model, lender data synced into an existing system and an enterprise platform, each quote linked to its project.

What RAG Development Covers

Retrieval-augmented generation answers a question from your own documents. It finds the passages that match, hands them to a model with the question and returns an answer that cites them. RAG development is building that pipeline, the index behind it and the tests around it, inside the systems your team already uses.

The goal

Answers your team can check in one click, from documents each person is allowed to open.

RAG, or Something Else?

RAG answers questions from documents. Some questions need a database query, a form reader or an agent, and we say so before anything is built.

RAG fits when

  • The answers live in documents that changePolicies, manuals, contracts and tickets that get new versions every month.
  • People need to see the sourceEvery answer has to point to the passage it came from, so a person can check it.
  • Access differs by userTwo people asking the same question may see different documents.

Which Documents Come First?

Tell us where the answers live today and who asks for them. We will say whether RAG fits and what a first index should cover.

We reply within 24 hours and can sign an NDA first.

Have files or a long brief? Use the full form

The Index Keeps Your Permissions

A RAG system must never show a passage the person asking could not open in the source system. Access rights go into the index with every chunk.

One chunk in the index

Passage
Contractor access to staging requires a signed access request.
Source
Access policy, section 4.2
Customer ID
••••••••
May open
EngineeringSecurity
  • An engineer asks

    Groups: Engineering

    The passage can be retrieved

  • A sales rep asks

    Groups: Sales

    Removed before the search runs

  1. 01Access groups on every chunk

    Each chunk carries the groups that may open its source document, copied from the system it came from and refreshed with every sync.

    A filter that trims results by group or user identity
  2. 02Filtered before the search

    The user’s groups filter the index before any search runs. A passage the user may not open never reaches the ranking, the prompt or the model.

    The filter is applied before the search engine runs the query
  3. 03Masked before indexing

  4. 04IDs instead of whole documents

    Prompts carry the passages and their source IDs, never full files. The answer links back to a document the user could already open.

Need the whole system reviewed? Security consulting covers threats and controls beyond the index.

Security consulting

Where the Index Lives

The vector store is a decision about your data, your scale and your cloud, not a default. We pick it with a test on your own questions.

  1. 01In the database you already run

    Choose when the records already live in PostgreSQL or Azure SQL and one database is simpler than two.

  2. 02Managed search on Azure

    Choose when you want keyword and vector search, reranking and security filters in one service.

  3. 03Dedicated vector databases

    Choose for very large collections, heavy filtering or a deployment outside Azure.

  4. 04Prototypes and local runs

    Choose to test retrieval on a laptop before the index has a home.

Around the index in our AI builds

  • Azure OpenAI
  • Azure AI Document Intelligence
  • .NET
  • PostgreSQL

Scored Before Every Release

A new model, prompt or chunk size can quietly make answers worse, so every release runs the same test set first.

The test set

Real questions from your team, with the answers and passages your experts expect

It grows with every question the system got wrong or could not answer.

Retrieval

Asks

Did the search find the passages an expert would use?

Fails a release when

The right passage is missing from the top results.

Groundedness

Asks

Does every claim in the answer come from those passages?

Fails a release when

The answer states something no passage supports.

Relevance

Asks

Does the answer address the question that was asked?

Fails a release when

It answers a nearby question, or buries the answer.

Completeness

Asks

Does it cover everything the expected answer covers?

Fails a release when

A condition, exception or step the expert listed is left out.

The release gate. All four scores hold on the test set, or the release waits while we find which step slipped.

The same control, in the invoice recognition SaaSOutput validated across templates and vendors, with corrections fed back into retraining

US Office, European Engineering

Headquartered in Warsaw, Poland, with a client office in New York. A European engineering team that works in your hours, with a US point of contact.

Headquarters

Warsaw, Poland

Engineering and delivery, and the main contact point for clients in Europe.

Client office

New York, United States

Kickoffs, contracts and account management for North American clients.

Shared hours with North America

Clients in Europe share the full working day with the team.

8 AM to 12 PM New York time, 2 PM to 6 PM in Warsaw.

Rankings for Our AI Team

Independent rankings for the engineers who build our retrieval and AI work, each next to the service it relates to.

Related reading

Read Before You Index

Guides from our team on OCR for invoices, choosing cloud storage and supervised and unsupervised learning.

All Insights

Frequently Asked Questions

Questions teams ask before they put an AI system in front of their own documents.

Have a different question? Talk to our team

Ready to Try RAG?

Tell us which documents, who asks the questions and the systems they sit in.

We reply within 24 hours, and we can sign an NDA before you share details.

Talk to the people who will scope your project