Services for businesses in Lucknow, India and beyond
Knowledge systems that answer from your documents and show the passage they used.
A knowledge system lets a team ask questions of documents it already has, and get an answer that points back to the passage it came from. The usual method is retrieval-augmented generation, or RAG. The software finds the relevant excerpts, and a language model writes an answer using those excerpts. The model is not retrained on your company. If the fact is not in the documents you provided, the honest result is that the system does not know.
Sunbeam builds document Q&A and larger knowledge systems from Lucknow for organisations in India and for clients in the Gulf, the UK and the US. Most of the work is the corpus and the permissions. The question box is the smaller part.
The business case
What better looks like.
- Answers a reader can trace to a source
- A collection with an owner and a current set
- A record of questions the documents could not answer
What it is
Document Q&A that shows its sources
Someone asks in the words they actually use. The system searches an index of approved documents, selects a handful of passages, and asks a model to answer from those passages, with a citation the reader can open. If the passages disagree, the answer should say so. If nothing relevant comes back, the answer should say that, rather than improvising.
Document control decides which file is current and who approved it. Q&A reads content that is already trustworthy. A pilot usually covers one team and one collection: a handbook, manuals, procedures, or support replies you are willing to reuse. A production system adds permissions, a way to replace a stale file, and a review step when a wrong answer would be expensive.
Answers with a trail
Every substantive answer cites the document and the passage, so the reader can check and the owner can see a weak source.
A corpus with an owner
Each collection has someone who can say whether a file is current. An index without an owner decays after launch.
Permissions copied from real access
If a contract is restricted to two people, retrieval must not hand it to everyone else because the question sounded related.
A refusal when the source is missing
The system is required to say when it cannot find a basis for the answer.
Who it is for
Teams that already wrote the answer down
This fits support teams, operations, professional services, and any group that inducts people by pointing at a shared drive. The value shows up when the same question is asked several times a week, the answer exists, and finding it takes a senior person's time. It fits poorly when the real policy lives in someone's head, when documents contradict each other and nobody will resolve them, or when a wrong answer would go to a customer without review.
Legal positions, clinical decisions and financial advice can be drafted from source text. They should be checked by a person before they are issued. We design that review queue, or we leave the use case out of the pilot.
Induction and procedures
New staff ask, then open the cited procedure. Seniors spend time on questions that are actually new.
Replies that must stay on policy
A draft grounded in the current article, which a person edits before it is sent.
A large manual
The paragraph that applies, for installers, operations staff or engineers who will not reread the whole PDF.
Files you may index
You have the right to process them, and you know which ones are confidential or out of bounds.
How we work
Decide the documents before you decide the model
We start with the questions people really ask and a sample of the files that should answer them. We look at format, duplicates, scans, and whether the important truth is in a table. Scanned pages with no text layer are called out. Some material needs cleanup. Some should stay out. A pilot can then use one shared collection for one team, with nothing sensitive uploaded. A wider release needs access control tied to the source, a process for replacing a document, and a record of questions that came back with no source.
Which questions are in scope?
We write a set of them and use that set as the evaluation. A rehearsed demo question is not an evaluation.
Which files are current?
The owner includes, excludes or labels. Drafts and superseded policies stay out, or they are labelled so the current file is preferred.
Who may ask about what?
If that map is unclear, the pilot stays inside one group that can already see every file in the set.
What happens on a bad answer?
The reader can flag it, and the owner can see the flag beside the source. Flags are how the corpus gets repaired.
Typical stack
Your files, a retrieval index, and an interface that cites
The usual shape is Next.js for the question screen, Postgres for users, collections and the question log, object storage on AWS for the files, and a retrieval index built from approved text. Go fits the worker that extracts text and refreshes a document when the owner replaces the file. The language model is called when someone asks. It is not where the documents are stored. A few hundred pages do not need the same machinery as a firm-wide archive, and we say so.
Interface
Next.js
Where a person asks, reads the answer and opens the citation, plus the view where an owner replaces a file.
Ingestion
Go
Workers that extract text from the file types you allow and rebuild an entry when a document changes. Extraction failures stay visible.
Records
Postgres
Users, membership, document status and the question log, including the questions the documents could not answer.
Files
AWS
Private storage for the source files, limited to the application and the people who administer it, with the retention rule you set.
Timeline ranges
A pilot is weeks. Permissions and freshness take longer.
The short range is honest when the documents are already digital text and one person can approve the set. Scans, a contested policy library, or a permission model that does not exist yet move the work into the longer range. We would rather spend the first week discovering that than index the wrong folder quickly.
1 to 2 weeks
Corpus review
Questions, file types, duplicates, sensitivity, and a recommendation to pilot, to clean up first, or to wait until the documents can support answers.
3 to 6 weeks
A document Q&A pilot
One team, one approved collection, citations and a feedback control. Enough to learn whether the answers help.
8 to 14 weeks
A production knowledge system
Permissions aligned with who can open each document, a way to update or retire files, and a wider evaluation set.
After launch
Keeping it true
Replacing stale sources and reading the questions that failed. Without that habit the system becomes a confident archive of last year.
Process
How a knowledge system goes into use
You can stop after the review or after the pilot. Both are useful. A production rollout is a separate decision, made with evidence from real questions.
01
Gather questions and files
The questions staff already ask, and the folders that should answer them. We sample. We do not ingest the whole drive on day one.
02
Mark what is allowed and current
An owner includes or excludes. Sensitive collections stay out of a pilot that has no permission model yet.
03
Ask the evaluation set
We run the agreed questions, read the citations, and fix gaps before anyone else sees the tool.
04
Open it to the pilot group
One team uses it on real work and flags bad answers while the context is fresh.
05
Decide on permissions
If the pilot helped, we scope access control and the update path. If it did not, we repair the corpus or stop.
06
Hand over the habit
Who replaces a file, who reads the failed questions, and where the system is hosted.
Who this is for
For teams ready to improve an important part of the business.
- Teams who induct people by pointing at a shared drive
- Support groups that must stay on a written policy
- Owners of a manual that is more accurate than memory
Frequently asked questions
Useful questions before you begin.
Do you train a model on our documents?
A standard pilot retrieves passages at question time and cites them. Training a private model on the archive is a different project, and it is rarely the right first step.
What if two documents disagree?
The answer should show the disagreement and the sources. Someone still has to decide which file is current. The software does not settle the policy.
Can the whole company use the pilot?
A pilot is one team that can already open every file in the collection. Wider access waits until permissions are part of the scope.
Which files can be indexed?
Digital text: office documents, HTML, and PDFs with a real text layer. Scans need a cleanup before they are worth indexing.
Related capabilities
One connected engineering team.
Automation and AI
Business automation, AI automation, AI agents and receptionist pilots for one workflow. Sunbeam is based in Lucknow and works across India and worldwide.
Starter pilots
Small fixed-scope pilots: booking, intake, lead follow-up, dashboards, QR ordering, document Q&A, an AI receptionist, or one AI agent. No prices listed.
Business systems
Business systems for portals, CRMs, internal tools and document control. Built in Lucknow for teams in India, the Gulf, the UK and the US.
Considering whether a custom approach is justified? Read our build-versus-buy decision guides, explore industry workflows, or review how ownership and handover work.
Start with the business problem, not the technology.
Tell us what is getting in the way. We will help identify a useful next step.