Web platformReal estate · CRE
Relm: multi-agent AI that finds hidden deal risk in commercial real estate
Minutesaddress to Intel Report
Enterprise LLM & RAG
Your knowledge. Your model. Your cloud.
Private LLM architectures, retrieval (RAG) systems and fine-tuned models that answer from your data, cite their sources and never leak it.
Cited
answers from your documents
Private
deployment in your cloud
Evaluated
accuracy measured before launch
What you getEverything to go from idea to production.
01
RAG pipelines
Ingestion, chunking and embeddings, then hybrid keyword and vector search with re-ranking, so each answer is built from the best passages and cites the document it came from.
02
Fine-tuning
Domain-tuned models for a fixed tone, output format or classification task, trained on examples you approve and added only where retrieval and prompting fall short.
03
Access control
Permissions synced from your identity provider and checked at retrieval time, so each person only gets answers from documents they could already open.
04
Document intelligence
OCR and layout-aware extraction for PDFs, scans, contracts, tables and spreadsheets, with the hardest files checked by hand before they enter the index.
05
Private deployment
Runs inside your AWS, Azure or GCP account in the region you choose, or on-prem, with model calls on endpoints that don’t train on your data.
06
Evaluation & monitoring
Accuracy, unsupported claims and latency scored on a fixed question set before launch, then on samples of live traffic after it.
How we deliver
- STEP 01
Audit
Data sources, access rules, use cases.
- STEP 02
Index
Ingest and embed your knowledge.
- STEP 03
Tune
Retrieval, prompts and evals.
- STEP 04
Deploy
Private, monitored, scaled.
Private ChatGPTA private ChatGPT for your company.
One chat assistant for your whole team. It answers from your own documents with citations and keeps every prompt and file inside infrastructure you control.
Option A
Managed API
Frontier models from Anthropic or OpenAI under business API terms that exclude training on your data, with zero retention where offered. Fastest to launch.
Option B
Your cloud’s model service
The same model families served inside your AWS Bedrock, Azure OpenAI or Google Cloud account, in the region you choose, under your existing cloud contract.
Option C
Self-hosted open model
Llama or another open-weight model on your own GPUs, in your cloud or on-prem. No external calls, at the cost of more operations work.
Data residency
Pick the region once. Documents, embeddings, chat history and logs stay in it, and sign-in goes through your identity provider, so people only get answers from files they can already open.
What drives the cost
Users and questions per day, the model tier, how many sources you connect, and whether you self-host. Self-hosting swaps per-question fees for a fixed GPU bill.
Our default: start on your cloud’s model service. Move to self-hosting only when policy forbids any external model call.
Release gateHow we know a RAG system is ready.
A RAG system is ready when it passes a test your team wrote before it was built. We agree that test in discovery and run it on every change after.
Step 1
Build the question set
Real questions from your team, each paired with the correct answer and the document that holds it. They come from search logs, support tickets and the people who answer these questions today.
Step 2
Score four numbers
Did the right passage come back (retrieval accuracy), is the answer right, does every claim point to a source (citation coverage), and how often does it state something no document supports.
Step 3
Gate the release
Each number gets a threshold you sign off. A new prompt, model or chunking setting ships only if it clears all four, so an upgrade can’t make answers worse without anyone noticing.
After launch
A sample of live questions is scored on a schedule you choose, and every wrong answer becomes a new test case.
Who owns what
Your subject experts own the answers in the question set. We own the pipeline that has to pass it.
Our rule: no RAG system goes live without a question set and a pass mark your team agreed to. What RAG costs to build and run
Stack by category
Retrieval
- Pinecone
- pgvector
- Weaviate
Models
- Llama
- Anthropic
- OpenAI
Cloud
- Azure
- AWS
Industries
Build or buyRAG, fine-tuning or an off-the-shelf chatbot?
Most teams need retrieval first. Fine-tuning and off-the-shelf tools each have a narrower job.
| Your situation | Our recommendation |
|---|---|
| Staff want writing help and Q&A over a few uploaded files | Buy an enterprise chatbot plan |
| Answers must cite your systems and respect permissions | Build a private RAG system |
| Facts change weekly: policies, prices, inventory | RAG, not fine-tuning |
| You need a fixed format, tone or classification | Add fine-tuning on top of RAG |
| Your documents are scans, tables and spreadsheets | Budget for document extraction before retrieval |
Our recommendation: buy for general help; build RAG when answers carry your data and your liability.
Typical engagementWhat a private RAG system costs and how long it takes.
- 01 · 2 weeks
Discovery Sprint
$5,000fixed fee · 2 weeksData sources and access rules audited, plus a prototype answering questions over a sample of your documents.
- 02 · 6–16 weeks
Fixed-Price Build
$25k–$120kper project · 6–16 weeksProduction RAG with connectors, hybrid search, permissions, evaluations and private deployment.
- 03 · Ongoing
Dedicated Team
From $12kper month · cancel anytimeNew sources, use cases and model upgrades, with accuracy tracked every sprint.
Built by one in-house team, from discovery to launch and support. AI makes us fast; a senior engineer decides what goes live. Meet the team
What moves the price: the number of sources, how messy the documents are, how strict the access rules are, and whether models run in your own account. Compare engagement models
Selected workShipped with this.
Web platformGuidesRead before you build.
- Buying software · 11 min readRAG development cost in 2026: what drives the price
- Engineering · 9 min readRAG vs. fine-tuning: choosing an enterprise LLM architecture
- AI strategy · 12 min readEU AI Act and GDPR for AI agents: a build checklist
- AI strategy · 11 min readBuild vs buy AI agents: a CTO decision framework
FAQCommon questions.
RAG or fine-tuning?
RAG for facts that change and need citations, such as policies, prices and product data, because you update the index instead of retraining a model. Fine-tuning for consistent behavior: a fixed output format, tone or classification. Most production systems start with RAG and add fine-tuning only where prompts can’t hold the format.
Will our data train public models?
No. We use endpoints that do not train on your data, with zero retention where the provider offers it, or self-hosted models, and your data stays in your environment.
Can it handle scanned documents?
Yes. OCR and layout-aware extraction turn scans and PDFs into searchable, structured data.
How much does a RAG or enterprise LLM system cost?
Most engagements start with a 2-week Discovery Sprint ($5,000 fixed fee, credited to your build), followed by a fixed-price build ($25k–$120k per project, 6–16 weeks) or a dedicated team (from $12k per month). The build price is fixed after discovery and billed in milestones tied to demoed features. Scanned PDFs, tables and spreadsheets take more ingestion work than clean text. Embedding, vector storage and model usage are separate running costs we estimate in discovery.
How long does a RAG system take to build?
Discovery takes two weeks and ends with a prototype answering questions over a sample of your documents. A production system with connectors, access control, evaluations and monitoring is a fixed-price build of 6–16 weeks.
Where does our data live?
In the cloud account and region you choose (AWS, Azure or GCP) or on-prem. Documents, embeddings, chat history and logs stay there. Model calls go to API endpoints that do not train on your data (zero retention where the provider offers it) or to a model hosted in your own account. For EU data residency we can deploy in an EU region such as Frankfurt or Ireland, with model calls on endpoints that keep processing in Europe, such as Vertex AI’s EU multi-region (for the models that support it) or Azure’s EU Data Zone.
Which vector database do you use?
Postgres with pgvector when you already run Postgres and want one less system; Pinecone or Weaviate when you need managed scale or advanced filtering. Retrieval combines keyword and vector search either way.
How do you measure answer quality?
We build an evaluation set of real questions with known answers and sources, then track retrieval accuracy, answer correctness, citation coverage and hallucination rate on every change, before and after launch.
What are the monthly running costs of a RAG system?
Usually far less than the build. In our worked example, 10,000 questions a month over 20,000 pages of documents comes to roughly $80 to $210 a month for model tokens, embeddings and the vector store at list prices, with hosting and monitoring on top. Self-hosting a model swaps those per-question fees for a fixed GPU bill.
Should we build a private LLM or buy an enterprise chatbot?
Buy an enterprise chatbot plan when general writing help over a few uploaded files is enough. Build when answers must respect per-user permissions, cite your systems of record, run in your own cloud, or feed other software through an API.
