AI / RAG System · 2025

Sauti — an AI assistant that answers from your documents, with citations.

A retrieval-augmented assistant built over private company knowledge — policies, contracts, PDFs and internal wikis. Every answer is grounded in a source the user can open, so trust is verifiable, not assumed.

Read the case study Build something similar
92%Answer accuracy on eval set
<2sMedian response time
−78%Repeat questions to support
sauti — knowledge assistant
What's our refund window for enterprise plans?
Sauti · grounded Enterprise plans carry a 30-day refund window from the invoice date. After that, unused seats can be credited to the next billing cycle.
Policy 4.2 MSA §7.1 Finance FAQ
Can we pro-rate mid-term?
The problem

Knowledge existed. Answers didn't.

The organisation had years of documentation spread across drives, inboxes and PDFs. Staff couldn't find what they needed, so they asked each other — and the same questions kept coming back to the same three people.

Before

  • Answers scattered across drives, email and chat
  • Senior staff interrupted constantly with repeat questions
  • Generic AI tools hallucinated policy details — unusable

After

  • One assistant covering every internal document source
  • Answers cite the exact clause, page or policy they came from
  • An eval harness that catches regressions before users do
Capabilities

What Sauti actually does.

Document ingestion

PDFs, DOCX, spreadsheets, web pages and wikis are parsed, cleaned and versioned automatically as sources change.

Hybrid retrieval

Vector search combined with keyword matching, so both semantic questions and exact clause references land correctly.

Citation grounding

Every response links back to its source chunk. If the answer isn't in the documents, Sauti says so instead of guessing.

Admin console

Non-technical staff can add sources, mark documents stale, review unanswered questions and tune tone.

Evaluation harness

A labelled question set runs on every change, scoring accuracy and retrieval hit-rate so quality is measured, not felt.

Multi-channel

The same engine powers an internal web app, a WhatsApp bot and a public help widget — one knowledge base, many doors.

How it works

The RAG pipeline, end to end.

01

Ingest

Sources are fetched on a schedule, parsed, deduplicated and assigned stable IDs.

02

Chunk & embed

Documents are split on semantic boundaries and embedded into a vector store with metadata.

03

Retrieve

A hybrid query pulls the top candidates, reranks them, and drops anything below a confidence floor.

04

Answer with sources

The model responds only from retrieved context, attaching citations the UI renders as links.

Results

Measured, not claimed.

92%Accuracy across a 180-question labelled evaluation set
−78%Drop in repeat internal questions routed to senior staff
1.8sMedian end-to-end response time including retrieval

“It gives the same answer every time, and it shows you where it came from. That's the part that made people actually trust it.”

— Operations lead, pilot deployment
Stack

Built with.

PythonFastAPI LangChainLlamaIndex pgvectorPostgreSQL OpenAI APIRedis ReactDocker AWS

Need an assistant that actually knows your business?

I build RAG systems that stay grounded in your data — with citations, evaluations and an admin console your team can actually use.

Start a conversation Back to portfolio