ChatGPT and Gemini were trained on the public internet. They do not know your business. Codeora Vision builds private LLM knowledge systems: RAG pipelines on LangGraph, MCP, and Claude that ground every answer in your documentation, your catalog, your clinical guidelines. Self-hosted or VPC-deployed, HIPAA and SOC 2 ready, citations included, agent-native, not retrofitted. Your data stays yours.
Trusted by teams in healthcare, legal, and finance handling regulated, sensitive data.
Updated June 7, 2026
WHAT IT IS
A private LLM is a large language model deployed so your data never leaves your control. It runs self-hosted, on-premise, or in your private cloud VPC, usually paired with RAG so it answers from your own documents and cites them. Unlike a public chatbot, nothing you send to it trains someone else's model.
First-mention enrichments: private LLM (a model you run in an environment you control, not a shared public API) · RAG (Retrieval-Augmented Generation, retrieval that grounds answers in your documents) · VPC (virtual private cloud, your isolated cloud network) · grounding (tying every answer to a real source you own) · air-gapped (a system with no outside network connection at all).
Per McKinsey's Superagency report, 92% of companies plan to invest in AI, yet only 1% consider themselves mature.
Maturity is hard partly because public models do not know your business and cannot touch your private data. A private LLM grounded in your own knowledge is how that gap closes, and your data stays yours.
THE PROBLEM
A public model has three problems for serious work. It hallucinates on your specifics, because it never saw your data. It cannot answer from your documents at all without help. And sending your records to a third-party API risks data leakage, vendor lock-in, and token-based pricing surprises. For regulated teams, that combination is a non-starter.
HOW IT WORKS
RAG works in four stages: chunk, embed, retrieve, and generate with citation. Codeora Vision indexes your documents in a vector database, then at question time retrieves the right passages and reasons over them with Claude or an open-weights model, orchestrated on LangGraph and MCP. Every answer ships with its source. Citations included.
We split your documents, catalog, or case files into clean, retrievable passages, tuned to your content structure.
We generate embeddings and index them in a vector database: Pinecone, Weaviate, Qdrant, Supabase pgvector, or Chroma, picked for your scale and privacy needs.
On each question, the system runs semantic and hybrid search, then re-ranks the matches so the most relevant, on-source passages win.
LangGraph, MCP, and Claude (or a self-hosted Llama model) generate the grounded answer with attribution, so hallucination prevention is built in, not hoped for.
This is the whole point. An answer you cannot trace is a guess. Every answer our private LLM gives links back to the source passage, so a clinician, lawyer, or analyst can verify it in one click.
GROUND YOUR AI
Bring a sample of your documentation, clinical guidelines, or case files. In a 30-minute review we will scope the private LLM and RAG system that grounds answers in it, the deployment that fits your compliance posture, and the citation model. NDA from minute one.
WHERE IT RUNS
The deployment is the wedge. Most vendors give you one option, their cloud. We support the full range and pick the one your data sensitivity and compliance posture require.
The model and data run on your own hardware, open-weights through vLLM or Ollama.
Best for: Maximum control, internal infrastructure teams
Deployed in your isolated VPC on AWS Bedrock, Azure OpenAI, or GCP Vertex AI, with a zero-retention endpoint.
Best for: Cloud-native teams who want privacy without managing hardware
No outside network connection at all, open-weights only, for the most sensitive data.
Best for: Defense, healthcare, and finance with the strictest rules
Under the hood we run Docker and Kubernetes for orchestration, vLLM and Ollama for self-hosted serving, and LiteLLM to keep model routing flexible, so the stack is portable, not locked to one vendor.
COMPLIANCE
This is the section sophisticated buyers actually verify, especially in healthcare and legal. We scope the controls to your regulatory posture before a single document is indexed, not after. Your data stays yours.
HIPAA
PHI-safe retrieval for clinical and medical records, with BAAs signed before launch.
SOC 2 Type II
Builds designed for SOC 2 controls: access logging, audit trails, and change management.
GDPR · CCPA
EU, UK, and California data rights, retention limits, residency, and deletion on request.
ISO 27001
Information-security controls aligned for enterprise procurement review.
Encryption
Encryption at rest and in transit on every deployment, with zero retention on private endpoints.
Attorney-client privilege
Privilege-preserving retrieval and access controls for legal case files and contracts.
In an air-gapped or on-premise build, your documents and your model never touch the public internet. Nothing trains a vendor's model. Nothing leaves your boundary. Your data stays yours, in the literal sense.
WHERE IT WORKS
RAG is the strongest cross-vertical capability we build. The pipeline is the same. The knowledge base, the compliance scope, and the source type change per industry.
Clinical guidelines and medical records retrieval, HIPAA-scoped and BAA-backed.
Case law search and contract analysis, privilege-preserving, where we go beyond point tools like Harvey AI by building on your own corpus.
A product knowledge base for support and merchandising, grounded in your catalog.
Internal wiki, employee handbook, and technical documentation search across your tools.
When the front end is a conversational assistant, the same retrieval layer powers our chatbot knowledge layer. This page is the knowledge system. That page is the chatbot.
OUR STACK
This is where RAG buyers test whether an agency actually builds. We are model- and database-agnostic and choose per project, on scale, privacy, and accuracy.
VECTOR DATABASES
PROPRIETARY MODELS
OPEN-WEIGHTS (SELF-HOSTED / AIR-GAPPED)
RAG FRAMEWORKS
SERVING & INFRA
EVAL & GOVERNANCE
Retrieval quality is where most RAG builds fail. We tune chunking, embeddings, hybrid search, and re-ranking against your real questions, and measure the result with LangSmith, so accuracy is proven, not assumed.
TRANSPARENT PRICING
We quote fixed-price by scope, not by tokens, so there is no metered surprise after launch. Projects start at $25,000. The build runs in a pilot phase, then a production deployment.
One private RAG knowledge base, one source, VPC or self-hosted, citations
OPTIONAL $2,000/MO
Scope this buildMulti-source RAG, private LLM endpoint, evaluation, observability, compliance controls
OPTIONAL $3,500/MO
Scope this buildOn-premise or air-gapped open-weights deployment, multi-source RAG, full governance
OPTIONAL $5,000/MO
Talk to an architectThe retainer is optional, not a lock-in. You own the system and the model weights you self-host. We stay on only for continuous evaluation and updates if that is worth it to you.
START
You do not have to commit to an air-gapped enterprise build to start. We often begin with a single-source RAG pilot in your VPC, prove the retrieval accuracy and citations on your real questions, then scale to production.
PROOF
CONTEXT
A national healthcare network needing clinical knowledge access without exposing PHI
WHAT WE BUILT
An air-gapped private LLM on Llama 3.3, RAG over clinical guidelines indexed in Qdrant, HIPAA-scoped with BAAs and audit logs
OUTCOME
Clinicians get grounded, cited answers from approved guidelines, with no data leaving the network
CONTEXT
A mid-market litigation firm searching case law and contracts by hand
WHAT WE BUILT
A private RAG system on Claude and Pinecone over their own case files, privilege-preserving, with every answer citing the source document
OUTCOME
Document retrieval that took hours now returns cited passages in seconds, inside their boundary
CONTEXT
An enterprise drowning in scattered internal documentation
WHAT WE BUILT
A self-hosted knowledge system on Mistral Large and Supabase pgvector, RAG over the wiki, handbook, and technical docs, served through their tools
OUTCOME
Staff get one grounded, cited answer instead of searching five systems, with the model running in their own cloud
Each outcome is a single anonymized engagement and not a guarantee. Architecture and entities are real; identifying details are not.
FAQ
A private LLM is a large language model your organization runs in an environment you control, so your prompts and documents never leave it or train a vendor's model. It can be an open-weights model like Llama 3.3 or Mistral self-hosted, or a model like Claude reached through a zero-retention endpoint on AWS Bedrock. Paired with RAG, it answers from your documentation and cites the source. We build them self-hosted, air-gapped, or VPC-deployed.
They solve different problems and usually ship together. A private LLM is about where the model runs and who controls the data. RAG is about how the model accesses your knowledge: it retrieves passages from your documents in a vector database like Pinecone or Weaviate and grounds the answer, with a citation. You can run a private LLM with no RAG, or RAG over a public API. We build both as one system.
Projects start at $25,000 and run to $80,000 or more, plus an optional $2,000 to $5,000 per month retainer. A single-source pilot in a VPC sits at the floor. An air-gapped enterprise system with multiple sources and full governance sits at the top. We quote fixed-price by scope, not by token usage, so there is no metered surprise after launch.
Both. We deploy self-hosted on your own hardware, fully air-gapped for the most sensitive data, or in a private cloud VPC on AWS Bedrock, Azure OpenAI, or GCP Vertex AI. Air-gapped builds run open-weights models like Llama and Mistral through vLLM or Ollama. Private-cloud builds can use Claude or GPT-4o through a zero-retention endpoint. We scope it to your compliance posture.
It can be built to meet all three, which is why regulated teams choose private. We design healthcare builds for HIPAA with signed BAAs and PHI-safe retrieval, build for SOC 2 Type II with audit logs, and handle GDPR and CCPA rights including deletion and residency. Encryption at rest and in transit is standard, and air-gapped builds keep data inside your boundary entirely.
A single-source RAG pilot typically goes live in three to five weeks. A full production private LLM with multiple sources, evaluation, and compliance controls runs six to ten weeks. Air-gapped on-premise builds run a little longer for infrastructure. Our Claude and LangGraph stack moves faster than from scratch, because the orchestration, retrieval, and evaluation patterns are proven.
For self-hosted or air-gapped builds, we use open-weights models: Llama 3.3, Llama 3.1 70B, Mistral Large, Phi-4, Gemma, DeepSeek, or Qwen, served with vLLM or Ollama. For private-cloud builds, we use Claude or GPT-4o through a zero-retention endpoint. We pick on accuracy, cost, latency, and how private the deployment must be, not on a default.
ChatGPT Enterprise promises not to train on your data and adds admin controls, but your prompts and documents still travel to OpenAI's servers through a third-party API. A truly private LLM runs in your own VPC, on-premise, or air-gapped, or uses a zero-retention endpoint, so data never leaves your boundary. For HIPAA, attorney-client privilege, or air-gapped needs, that difference is the whole point.
DALLAS, TEXAS — RESPONDING IN < 4 HOURS
Send us a sample of your documents, guidelines, or case files. We will scope a private LLM and RAG system on Claude or open-weights models, deployed where your compliance posture requires, with citations on every answer. You own it. Your data stays yours.