Skip to content
Valtair
Work
Revenue operationsB2B SaaS platformConfidential

A grounded knowledge assistant for a revenue platform

We built a knowledge assistant that answers questions across a revenue platform's own content, with citations and permissions, so teams stop hunting through scattered sources.

Illustrative example. Named engagements are shared with client permission.

01

Context

A growing B2B SaaS company had knowledge spread across documentation, support tickets, and product data. Teams wasted time searching, and answers were inconsistent.

02

Challenge

Support and sales teams could not find answers across scattered docs, tickets, and product data.

03

Product approach

We treated retrieval quality as the core problem. Content was ingested into a hybrid retrieval layer, answers were grounded in sources with citations, and access controls were enforced at retrieval so users only saw what they were permitted to.

04

System architecture

FrontendAn assistant embedded in the existing product surface.
BackendA retrieval API with permission enforcement and logging.
AIGrounded answer generation with citation checking.
RetrievalHybrid graph and vector retrieval over ingested content.
IntegrationsDocumentation, ticketing, and product data sources.
InfrastructureCloud deployment with monitoring on retrieval quality.
05

AI and data components

  • Grounded answer generation
  • Citation and support checking
  • Query understanding
  • Provider-agnostic model orchestration
06

Backend and integrations

  • Hybrid retrieval service
  • Permission enforcement at retrieval
  • Content ingestion pipelines
  • Answer and source logging
07

Engineering decisions

Hybrid retrieval over pure vector search

Some questions needed relationships between entities, so we combined graph and vector retrieval instead of relying on similarity alone.

Permissions at retrieval, not display

Access was enforced when retrieving content, so restricted material never reached the model in the first place.

08

Responsible AI and safeguards

  • Every answer carries citations to its sources
  • Permissions enforced so users only see what they may access
  • Answers checked for support against retrieved content
  • Retrieval quality monitored in production
09

Outcome

  • Faster answers for support and sales teams
  • More consistent responses grounded in approved sources
  • Less time lost searching across systems
10

Lessons

  • Grounding and citations are what make an assistant trusted enough to use.
  • Enforcing permissions at retrieval is simpler and safer than filtering later.
  • Retrieval quality, not the model, was the lever that mattered most.

Facing a problem like this?

Tell us what you are building. We will show you how we would approach it, and reply within one business day.