The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Enterprise RAG AI: Private Doc Intelligence
A fast-growing B2B software firm needed a secure, enterprise-grade Retrieval-Augmented Generation (RAG) system to query thousands of confidential internal documents (legal contracts, technical architecture specs, Notion pages, and employee SOPs) in real time without third-party data leakage.
Key Challenges:
High hallucination risk and irrelevant context retrieval using standard keyword search.
Managing multi-format files (scanned PDFs, tables, markdown) with strict access control based on user hierarchy.
Technical Implementation & Solution:
• Engineered an automated ingestion pipeline in Python with LlamaIndex and LangChain for semantic chunking and metadata preservation.
• Integrated a hybrid vector search architecture pairing dense embeddings (OpenAI text-embedding-3) with sparse BM25 indexing in Qdrant.
• Added a cross-encoder re-ranking layer (Cohere Re-ranker) to prune low-relevance chunks, reducing API token costs by 40%.
• Built robust Role-Based Access Control (RBAC) via FastAPI endpoints, ensuring departmental data isolation.
Results & Impact:
• Dropped internal information retrieval time from 20 minutes to under 4 seconds.
• Attained 98% factual precision with traceable in-line page citations.
• Full on-prem Docker deployment ensuring 100% private data security.
I'd test revoking a user's document access mid-conversation. In your RBAC flow, do cached retrieval results get rechecked against current permissions before they're sent to the reranker or model?
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
Rebuilding the moment an inbound inquiry turns into a confirmed booking for MEND Appliance Care, a local repair shop in Spalding, Lincolnshire.
The Problem Solved: Replaced 2.5 hours of daily owner phone-tag, vague fault descriptions, wrong van parts, and customer no-shows with an instant AI-powered diagnostic intake, fixed upfront pricing, deposit collection, and auto-generated dispatch scheduling.
Key Features Built in Lovable:
• Multi-Modal Intake: Customers describe the issue using text, photos, model numbers, or video.
• Instant AI Diagnosis & Transparent Pricing: Autodetects fault & exact part (e.g., Bosch Serie 6 drum shock absorbers) with clear labor/part cost breakdowns.
• Flexible Resolution Paths: Home Repair, Workshop Drop-Off, DIY Part Purchase, or local stockist routing (Toolstation/Screwfix).
• Owner Safety Net & Dispatch Dashboard: Real-time job calendar, "Ask Maya" human-in-the-loop review queue, and 1-click customer comms.
Built entirely with @Lovable for the #lovablechallenge.
Shop owner dashboard : Email: maya@mend-demo.co.uk
Password: MendDemo2026
The "Ask Maya" human-in-the-loop queue is one of those details that separates a clever demo from something you can actually ship to a real business. Most AI intake flows skip this — they either automate everything and break on edge cases, or hand off to a human and defeat the...
Problem: Apps need a secure, organized backend to handle users, orders, products and AI features.
Solution: Built a centralized API Gateway (BFF) in n8n Cloud. One webhook handles validation, API key security and rate limiting, then routes to sub-workflows: Users (register/login), Orders (stock check, totals, status), Products (search), and AI (OpenAI chat, recommendations, review summaries). Supabase is the database.
Tools: n8n, Supabase, OpenAI API, Webhooks
Self-initiated demo project.
I'd test authorization separately for each sub-workflow, especially Orders. Does the gateway pass a verified user identity through, or does the Orders workflow check that the caller can access each order?
Designed and built an AI operations dashboard that brings projects, decisions, tasks, research, connected tools, and voice interaction into one clear workspace. The goal is to turn complex workflows into a simple, usable system that helps people see what matters and act faster.
James, putting decisions right next to tasks is the part I like most. Most ops tools track what got done but lose why it was decided, and that is exactly what people go looking for three weeks later. How are you using the voice side day to day?