Back to projects

React, RAG, LLM streaming, vector retrieval

Internal AI knowledge assistant — DXS

Senior Frontend Developer · DXS

I helped bring AI into DXS consulting workflows with an internal knowledge assistant: documents are ingested and chunked, retrieved against a vector index, and answered through a streaming UI with source citations. The goal was daily use by consultants — not a demo chat — so grounding, error states, and citation UX mattered as much as the model call.

documents → ingest / chunk → embeddings → retrieve
     → LLM stream → answer UI + citations → consultant review

Deep dive

From prototype to daily tool

Internal AI only sticks if it survives real consulting pace. We moved from a sandbox chat to a surface with loading, retry, empty retrieval, and citation affordances so people could verify claims against the source pack before putting language in front of a client.

Streaming interface

Answers stream token-by-token with interruption and reconnect behaviour. Citations map back to retrieved chunks so the UI can show where a claim came from — critical when the audience is consultants who will not paste unchecked LLM text into deliverables.

Enablement loop

Alongside the product I trained 20+ consultants on practical AI use: when to trust retrieval, how to prompt for grounded answers, and where automation helps vs. where human judgment stays mandatory.

Highlights

Document ingestion, chunking, and vector retrieval for internal knowledge
Streaming answer UI with inline source citations and failure states
Built for consulting workflows — verify before client-facing use
Paired with hands-on AI enablement across the organisation