- Home
- Case Studies
- About
- ServicesSoftware Engineering
- Custom SaaS DevelopmentNewMulti-tenant, billing, SOC 2-ready, AI-native.
- Next.js DevelopmentProduction Next.js 15 with App Router, RSC & edge.
- Python & FastAPIHigh-performance APIs, microservices, AI inference.
- Mobile App DevelopmentNewReact Native & Flutter — iOS + Android.
- Web ApplicationsFull-stack web apps, portals & dashboards.
- Custom SoftwareBespoke software, sprint delivery, real PMs.
AI & Data- AI Agent DevelopmentLangChain, OpenAI Agents SDK, Claude — RAG + tools.
- Custom RAG DevelopmentNewHybrid search, reranking, evals & citations.
- MCP Server DevelopmentNewExpose your tools to Claude, Cursor, ChatGPT.
- AI & ML DevelopmentComputer vision, predictive analytics, fine-tuning.
17 years · 1000+ projects · Toronto HQ - Clients
- Career
- Contact
- Blog
Quick answer
UnlockLive IT designs and ships production retrieval-augmented generation systems — the kind that hold up in front of real users with real questions on real, messy enterprise data. We build hybrid retrieval pipelines (BM25 + dense embeddings + reranking), permission-aware search across your Notion, SharePoint, Confluence, S3, and Salesforce data, and answer generation with strict citation requirements and automated evals on every change. Default stack: Python/FastAPI + Qdrant or pgvector + Cohere Rerank + Claude or GPT-5, deployed on AWS, Modal, or fully on-prem for regulated workloads. See our Local LLM case study for a worked example of air-gapped RAG.
What we build
Enterprise knowledge-base search:Search-and-answer over your company's Notion, Confluence, Google Drive, SharePoint, Slack, Zendesk, Jira, and internal wikis — with permission-aware retrieval (a user only sees results from documents they're allowed to read).
Customer-facing support copilots:RAG over your help center, product docs, release notes, and historical tickets. Streaming answers with inline citations, deflection metrics, and clean handoff to a human agent when confidence drops.
Domain-specific Q&A on regulated content:Legal contracts, medical literature, clinical guidelines, financial disclosures, building codes — with strict citation requirements and full audit trails for compliance review.
Sales enablement and proposal copilots:RAG over win/loss notes, past proposals, pricing decks, ICP profiles, and competitor battle cards — generating first-draft RFP responses and discovery prep in seconds.
Code-aware RAG over your repos:AST-aware chunking of your codebase plus README, ADR, and ticket context — for internal developer copilots and onboarding assistants.
Multi-modal RAG (PDFs, images, tables, video):Layout-aware PDF parsing (Unstructured, LlamaParse, Reducto), table extraction, vision models for charts and screenshots, and video transcript indexing.
Our RAG technology stack
Our RAG development process
- Use case definition (1 week): Define WHO is asking, WHAT they're trying to accomplish, what an acceptable answer looks like, and the failure modes that would damage trust. We refuse to start RAG projects without an explicit success metric — usually answer accuracy on a labeled eval set.
- Data ingestion & cleaning (1-3 weeks): Connectors to source systems (Confluence, Notion, S3, SharePoint, Salesforce, Zendesk), document parsing, deduplication, ACL extraction, and metadata enrichment. Most retrieval problems are actually ingestion problems.
- Retrieval baseline & eval set (1-2 weeks): Hand-curate a 100-300 question eval set covering the long tail of real queries. Build a baseline retrieval pipeline. Score it. Document where it fails.
- Iterate on retrieval (2-4 weeks): Hybrid search, reranking, query rewriting, contextual retrieval, parent-document retrieval, metadata filters. Each iteration is measured against the eval set, not vibes.
- Generation & guardrails (1-2 weeks): Prompt engineering for the answer model, citation requirements, refusal handling for out-of-scope questions, jailbreak/prompt-injection defense, output classifiers.
- Production deployment & monitoring (1-2 weeks): API endpoints, caching, rate limiting, per-tenant isolation, full observability (LangFuse / Helicone), and dashboards tied to your business metric — deflection rate, time-to-answer, or analyst hours saved.
Frequently asked questions
What is RAG and why do I need a custom one?
Retrieval-augmented generation is the pattern where an LLM is given relevant documents from your knowledge base and asked to answer using only those documents. The off-the-shelf 'upload-PDFs-to-ChatGPT' versions work for demos but break in production because real enterprise data has access control, messy formats, custom chunking needs, citation requirements, and quality bars that demand iteration. Custom RAG is what you build when an answer being wrong has a real business cost.
How much does it cost to build a production RAG system?
A focused single-source RAG (one knowledge base, one user surface, English only) typically ranges from $30,000 to $80,000. Multi-source enterprise RAG with permission-aware retrieval and connectors to 5+ systems ranges from $80,000 to $200,000. Regulated-industry RAG with audit trails, on-prem deployment, and rigorous evals starts at $150,000. Inference and embedding costs are separate and depend on document volume and query rate.
Pinecone, Weaviate, Qdrant, or pgvector?
pgvector is our default if you already have PostgreSQL — for under ~10M vectors and moderate QPS, it's fast enough and removes a moving part. Qdrant is our default for self-hosted or air-gapped deployments. Pinecone or Turbopuffer are our defaults for very high-scale managed services where you want zero ops. Weaviate is a good choice when you want hybrid search out of the box and already have the operational appetite. We pick after benchmarking against your data.
How do you handle hallucinations and accuracy?
Three layers. (1) Retrieval quality — measured against a labeled eval set on every change, with reranking, hybrid search, and contextual retrieval to push recall higher. (2) Generation guardrails — prompts that require citation of retrieved chunks, refusal of out-of-scope questions, and a human-in-the-loop path for low-confidence answers. (3) Production monitoring — sampled human review of conversations, automated factuality checks, and immediate alerting on failure-mode regressions. Every system we ship has documented acceptable failure modes.
Can the RAG respect user permissions?
Yes. We extract ACLs from source systems at ingestion time, store them as filterable metadata in the vector database, and apply per-user filters at query time. For high-trust environments we also do post-retrieval permission verification against the source system as a defense-in-depth check. Permission-aware retrieval is non-negotiable for any deployment touching internal documents.
Can we deploy this on-prem or in our own VPC?
Yes. We routinely deploy entirely inside customer AWS / Azure / GCP accounts, and for regulated workloads we deploy fully on-prem and air-gapped using open-source models (Llama 3.3, Qwen, Mistral) served on vLLM or TGI, plus Qdrant for retrieval. See our Local LLM case study for a worked example.
How is RAG different from fine-tuning?
RAG retrieves fresh facts at query time and is the right answer for changing knowledge bases, citation requirements, and access-controlled content. Fine-tuning bakes patterns and style into the model and is the right answer for teaching the model a domain-specific format, terminology, or persona. They're complementary — many production systems do both. We help pick the right tool in the discovery phase.
Ready to ship a RAG that actually answers correctly?
Tell us about the questions you want answered and the data you want answered from. We'll respond within one business day. Book a free strategy call with our Toronto team.
UnlockLive IT Limited provides state-of-the-art software solutions to clients all over the world. We are a creative force with strategic alliances distributed around the globe
Copyright© UnlockLive .All right reserved 2008 - 2026