Broadcaster
A broadcaster whose staff spend measurable time finding internal guidance that already exists in documentation
AI & automationReady to pilotBuilt by LabsLast updated 29 July 2026
AI & automationReady to pilotBuilt by LabsLast updated 29 July 2026
Custom Knowledge Assistant
Answers drawn from your own documents, with access control enforced
Staff waste hours searching internal documentation for answers that already exist - and a general-purpose AI assistant cannot be trusted because it would show people content they are not cleared to see. Custom Knowledge Assistant answers directly from your organisation's documents, enforces per-user access control, and combines hybrid retrieval with a stateful agent so the answers are grounded and the access is governed.
LangGraph Stateful AgentHybrid RAG RetrievalAWS Bedrock Intelligence
Try the demo
Ask a question about Big Blue Marble
Answers come from a knowledge base built from publicly available Big Blue Marble material, with a source citation on every claim.
The problem this solves
Knowledge is locked inside documents, intranets and wikis that are too large to search manually and too sensitive to hand to an unrestricted AI. Teams resort to asking colleagues, who either know from memory or have to search themselves - the same cost, paid again and again by different people.
A general-purpose assistant solves the retrieval problem but creates a worse one: it cannot distinguish between a document that is public and one that is restricted, so deploying it exposes protected material to whoever asks the right question. The only safe answer is a system that retrieves and answers within the boundaries of what each user is actually allowed to read.
Who this is for
This is relevant if you are:
Telco
A telco with layered access policies who needs an AI assistant that respects those policies rather than overriding them
How it works
- 1
Documents are ingested and indexed twice: FAISS dense vector embeddings and BM25 keyword search, so retrieval works on meaning and on exact terms simultaneously. The indexes are synchronised from S3, so an updated corpus reaches the assistant without a redeploy.
- 2
When a query arrives, the LangGraph agent decides which tools to call - shortlisting the likely source documents, then retrieving from the knowledge base - and loops until it has enough to answer. Which tools exist at all is fixed per deployment, so an assistant built for one corpus cannot reach another one's sources.
- 3
Reciprocal Rank Fusion merges the ranked results from both indexes, so the most relevant passages rise regardless of which retrieval method found them.
- 4
Access control is applied at retrieval time: unauthenticated users reach only public collections; authenticated users reach the private sources their profile permits.
- 5
A separate zero-temperature LLM pass reviews the response and emits structured feedback metrics, creating a quality signal that improves the system over time.
What makes it different
Most enterprise knowledge tools are either open - so they ignore access policy - or siloed per team, so the person who needs an answer from a different department cannot get it. Custom Knowledge Assistant enforces the access boundary at retrieval rather than at the UI, which means the model never sees what the user is not allowed to see. Hybrid retrieval with RRF means it does not choose between keyword precision and semantic recall: it uses both and lets the ranking decide.
Current status & next steps
Ready to pilotReviewed 29 July 2026
Custom Knowledge Assistant is ready to pilot. The demo on this page is open, and deliberately narrow: its knowledge base holds only publicly available Big Blue Marble material, so it can show how the assistant retrieves and cites without putting anything private behind a public URL.
Two boundaries hold in the current deployment and are worth knowing before a pilot is scoped. It has no internet access, so the agent cannot run live web searches or pull real-time external data, and the architecture is text-only: it does not embed or reason over images, which keeps retrieval to text documents and metadata.
The next step is to run the assistant against a customer's actual document corpus and access policy, so quality and boundary enforcement can be measured on real material.
Want to deploy this on your own knowledge base?
Tell us what you need to keep private, what access structure you have, and we will tell you what a pilot against your corpus would involve.