AI agent · Knowledge
A network of real-estate advisors had its whole method in one large knowledge base. We built a coaching agent that finds the few sections a question needs and answers from them, instead of sending the entire base to a model on every call.
The situation
The network's know-how filled a knowledge base of roughly 196,000 tokens. Sending all of it with every question is slow, costly and still loses the right paragraph in the noise. New advisors needed answers in seconds, in the network's own words.
What changed
Each question now pulls only the relevant sections. Two search methods are combined so both meaning and exact terms are found, and a guardrail stops the agent from naming competitors or internal people while it streams its answer.
How it works
A coaching question from inside the network's platform.
Semantic search and keyword search run side by side over 962 sections.
Both result lists are merged into one ranking of the best passages.
The model writes a coaching answer from the selected sections only.
Names that must not appear are blocked as the answer streams out.
Where the AI stops
The model
The code
By the numbers
Built with
Built for
Real estateSaafkaA network of real-estate advisors: buy, sell, rent and appraise anywhere in Morocco.Read the client storyMore use cases
Contact
Write to usWe answer the same working day.