AI & Knowledge

AI Knowledge Base

Ask what your organisation already knows, and get the answer with its sources.

Bonyan AI Knowledge Base brings a company's email and attachments, documents, team chat, meetings and operational reports into one searchable memory. Ask in Persian or English and get a written answer with numbered citations to the exact message, file and page, built only from what you are allowed to read.

It runs on your own servers, from capture and indexing to search, AI answers and administration. It is built for operators, MVNOs and telecom companies that need their knowledge kept in-house.

Request a briefing

What did we agree with Partner A about the new interconnect rate?
Answer

The revised voice termination rate applies from the start of next month, after Partner A accepted the updated annex1. The billing team was asked to update the rate table before the next settlement run2, and the change was confirmed in the weekly wholesale review3.

Sources
  1. 1Re: Interconnect annex, revised ratesCommercial team · attachment Annex-2.pdf, page 3
  2. 2#wholesale: rate table updateTeam chat · two days later
  3. 3Weekly wholesale reviewMeeting transcript · at 14:20
Illustrative example with invented data. Every answer links back to the message, file and page it used.
At a glance

Built for Teams That Work in Two Languages

Persian and English questions, sources and meeting transcripts
فاEN
Every answer cites the message, file and page it used
[1]
Typical time for a weekly topic summary
<10s
Kinds of source: mail, documents, chat, meetings and reports
5
Data sources

One Memory for What the Company Writes and Says

Capture runs in the background. Nothing changes for staff: they keep using mail, chat and meetings as before.

Email & Attachments

All company mail captured on the mail server, with every attachment kept. Mailboxes are only ever read.

Documents & Scans

Uploaded documents, PDFs and spreadsheets; scanned pages and images are read with OCR.

Team Chat

Channel messages and shared files, under the same channel permissions.

Meetings

Recorded meetings transcribed in Persian on your servers, with speakers separated and a short summary.

Operational Reports

Daily batch and traffic reports from billing and mediation, turned into exact, queryable records.

Ingestion pipeline

From Raw Content to Trusted Knowledge

Every source goes through the same pipeline, so an e-mail, a contract scan and a meeting are searched, cited and governed alike.

  1. Capture

    Mail, chat, uploads and recordings are captured as they arrive, each with the people and groups allowed to read it.

  2. Extract

    Text, tables and attachments are extracted; OCR reads scans and images; speech recognition transcribes audio.

  3. Enrich

    Language normalisation, de-duplication, passage splitting, and tags for keywords, people, organisations and dates.

  4. Index

    Keyword and semantic indexes side by side, plus structured records for operational reports.

  5. Always on

    Govern

    Permissions, retention, archive and audit applied to every item and every answer.

Capabilities

Why Teams Trust Its Answers

Cited, permission-aware answers in both languages, with exact numbers where numbers matter.

01 · Cited answers

Answers You Can Check

Ask in plain Persian or English. The answer is written only from the passages found for you, and every statement carries a numbered citation to its source.

  • Each source shows the sender, date, subject, attachment and page
  • Deep mode reads the underlying messages and files in full for detailed questions
  • Retrieved content is treated as data: a document cannot change the assistant's instructions
Ask · one questionFrom question to cited answer
  1. UnderstandKeywords, dates, people, ids
  2. RetrieveKeyword + semantic, reranked
  3. FilterOnly what you may read
  4. WriteAnswer with numbered sources
Need more?Deep mode reads sources in full
03 · Permission-aware

Everyone Sees Only What They May Read

Access is decided per person and per item: email follows who was on the message; documents and chat follow directory groups and channels. The same question can return different sources for different people.

  • Enforced in the database on every query, not filtered afterwards
  • Staff sign in with their directory account; groups decide document and channel access
  • A topic summary is shown only to people who may read every item behind it
Same question · two people
  • Commercial manager7 sources used
  • Support engineer2 sources used
  • Messages the engineer was not onNever shown
Enforced in the database · on every query, every answer, every summary
04 · Topic chronicle

What Happened This Week, Without Rereading It

Name a topic, such as a project, customer, technology or event, with the other names people use for it. Each day gets one dated entry written from that day's mail, chat, files and transcripts, linked to every item it read.

  • Weekly and monthly questions are answered from the daily entries, typically in under 10 seconds
  • Ask for detail and the answer goes back to the original items
  • Quiet days stay quiet: nothing is invented to fill a gap
Topic · Interconnect project
  1. Revised annex sent to Partner A for review.
  2. No activity on this topic.
  3. Partner A accepted the new rate; finance informed.
  4. Rate table updated ahead of the settlement run.
Weekly summary · built from seven daily entries, each linked to its sources
05 · Exact report numbers

Numbers from Reports, Not Guesses

Daily reports that billing and mediation platforms send by e-mail are parsed into structured records, not just indexed as text. Questions about them are answered through reviewed forms that return exact tables.

  • The AI picks a form and fills in its parameters; it never writes a query
  • Corrections keep the earlier version, so history is preserved
  • Read more in Operational Report Intelligence
Report form · pricing batches, last 4 days
DayBatchesFiles rejectedErrors
Sat60212
Sun61198
Mon641,904
Tue60205
Exact table from a reviewed form · illustrative figures
06 · In AI assistants

Company Knowledge Inside Your AI Assistants

The knowledge base, the topic chronicle and the report forms are published as tools over the Model Context Protocol, so AI assistants can use company knowledge while you work.

  • Each person connects with their own key and their own permissions
  • The same cited answers as the web console
  • Usage is logged per person
AI assistant · company tools
Ask the knowledge baseList tracked topicsRun a report form
  • Signed in asYour own key
  • SeesOnly what you may read
Model Context Protocol · works in desktop and coding assistants
07 · On-premises

On Your Servers, with the Models You Choose

The platform runs in containers on your own virtual machines, with directory sign-in. Each AI job is assigned to a local model, a model server on your network or a cloud provider.

  • Original files, the index and the database stay in-house
  • If a cloud model is chosen for a job, only the passages that job needs are sent to it
  • Fully local operation is possible with local or in-network models
  • Meeting audio is always transcribed on your own servers
Models · one per job
  • Search index (embeddings)This server
  • OCR for scanned pagesModel server on your network
  • Answer writingLocal or cloud, your choice
  • Meeting transcriptionAlways on your servers
Changed in the console · no redeploy, one-click rollback for a new index
Security & governance

Built to Be Trusted with Sensitive Knowledge

Company mail is the most sensitive data most organisations hold. The platform is designed around that.

  • Read-only access to mailboxes: nothing is ever changed or sent
  • Permission filtering on every query, answer and summary
  • Ingested content is treated as untrusted data and cannot change the assistant's instructions
  • Retention, archive and permanent deletion under administrator control
  • Secrets encrypted and kept out of logs
  • Audit trail of administrative actions and data reads
  • AI usage and cost recorded per person
Administration

One Console to Run It

01

Pipeline & Jobs

Watch capture, extraction, OCR, transcription and indexing as they run.

02

Playground & Traces

Ask a question and see each step: what was understood, retrieved, filtered and cited.

03

Retrieval Tests & Evaluations

Check search quality on fixed questions before and after every change.

04

Models per Job

Assign embeddings, reranking, OCR and answer writing to the model each job should use.

05

Usage & Cost

AI use and cost per person and per day.

06

Archive, Restore, Purge

Remove an item from search and answers, restore it, or delete it permanently.

Use cases

Where It Helps Most

New Joiners

Learn how past projects, decisions and customer issues were handled, from the people who handled them.

Support Engineers

Find how a similar problem was solved before, with the mail and chat behind the fix.

Billing and Mediation Operations

Daily batch reports as exact tables, with unusual days easy to spot.

Wholesale and Interconnect

Agreements, rate changes and disputes followed day by day as a topic.

Meetings in Persian

Recorded meetings become searchable transcripts with a short summary.

Management Briefings

Weekly summaries of key projects and customers, linked back to the evidence.

Reference architecture

The Platform, Tier by Tier

Mail, chat, uploads and recordings are captured into an ingestion queue, processed into passages, indexes and records, and kept in one knowledge store where every row carries its permissions. The serving tier checks those permissions on every request before it searches, answers or runs a report form. AI models are assigned per job, and identity, security and monitoring run across every tier. The amber line is the path of a person's question.

AI Knowledge Base reference topology: email, documents, scans, team chat, meetings and operational reports are captured into an ingestion queue; a processing tier extracts, reads scans, transcribes audio, normalises Persian and English, splits, tags, embeds and parses reports; a knowledge store keeps text, permissions, keyword and vector indexes, records, topic entries, original files and the audit log; a serving tier filters by permission and provides search, cited answers, topic summaries, report forms, MCP and REST APIs and the web console to staff, AI assistants, internal tools, team chat and administrators; AI models run locally, in the network or in the cloud; directory sign-in, TLS, encrypted secrets, monitoring and containers span every tier
Open the full-size topology (opens in a new tab)
Why it matters

Benefits

Find It in Seconds

Past decisions and details without searching through mailboxes.

Check Every Answer

Numbered citations lead straight to the source.

Keep It In-House

Originals, index and database stay on your servers.

Persian First

Persian search and meeting transcription, alongside English.

Talk to us

See It Answer Your Questions

A briefing walks through cited answers, topic summaries and report forms, and how the system would run on your servers.

Request a briefing

Or e-mail info@bonyansystem.com.