The situation
Teams were losing days hunting for answers spread across Slack threads, Jira tickets and long documents. One client also needed accurate question answering in Arabic over a large religious text corpus, and a trading group wanted analysts to question two decades of commodity pricing without writing queries.
What we built
A unified operational assistant. A redesigned RAG system that brings Slack and Jira knowledge, and in a related build Salesforce, into one assistant. It supports cloud and on-premise LLM deployment, using LLaMA 3 and Mistral served with Ollama where data had to stay in-house.
Arabic extractive question answering. An extractive QA system built on open-source language models and vector search, returning precise spans from large Arabic corpora rather than generated paraphrase, which is what accuracy on religious texts requires.
Conversational querying over pricing history. Twenty years of commodity pricing modelled in a knowledge graph, so analysts can ask questions in plain language and get answers grounded in the stored series.
Results
- Internal time-to-answer on operational questions reduced from about two days to seconds
- About 90 percent answer accuracy on Arabic extractive QA
- Twenty years of pricing data queryable in plain language
Why it matters
Extractive answers, citations and on-premise options are what make retrieval usable where precision and confidentiality matter. The same patterns carry directly into regulated healthcare and financial knowledge bases.