Sosh customer support assistant - LLM + RAG

Context

Built for Sosh via Semiologic, this project set out to improve customer support. The idea was to use GraphComment as the answer module inside the interface, powered by an LLM + RAG layer to deliver clear answers to user questions.

This page documents the concept, the model review and the three flows I prototyped.

Challenge

• Bring generative answers into a customer-support interface without losing trust: answers had to rely on Sosh’s internal documentation, not the open web.

• Integrate the assistant into the existing help journey and the GraphComment answer module.

• Cover the real use cases: qualified answers, vague questions that need guidance, drafting help, and more.

Approach

• Benchmarked the main LLMs of late 2024 and early 2025, comparing UI/UX, answer quality, document grounding and cost on support questions.

• Designed an interface with an LLM + RAG layer connected to Sosh’s internal docs, exposed through GraphComment as the answer module.

Three prototype flows

• Qualified query: the user knows what they are looking for, the assistant returns a precise, sourced answer.

• Guided search: the user is not sure, the assistant accompanies them and narrows the request until it lands on the right answer.

• Tutorial: the assistant answers, then drafts a ready-to-publish forum post, so the answer serves the next person searching the same thing and keeps the forum alive.