Skip to main content
Enterprise RAG on AWS

Private AI that answers questions about your company's documents.

No third-party training. Your data stays where it already lives.

Shared multi-tenant on AWS, dedicated deployment in your own AWS account, or federated Connector for Azure, GCP, and on-prem. Built for multinationals, regulated industries, and mid-sized organizations that cannot afford for their knowledge to leave the perimeter.

Request early accessTalk to us

The problem

Critical knowledge is scattered

Methodologies, manuals, and processes live across SharePoint, Drive, Confluence, PDFs, and people's heads. New hires take months to onboard.

ChatGPT is not an option

Customer data, contracts, regulated information (GDPR, HIPAA-adjacent, finance). Uploading to public SaaS is a legal and reputational risk.

No traceability, no compliance

Without verifiable citations and forensic audit, no recommendation passes legal, risk, or internal compliance review.

What we offer

A complete, governed, enterprise-ready RAG platform

We don't sell a chatbot. We sell the full pipeline that turns your private knowledge into audited, evidence-backed answers — inside your own AWS infrastructure.

01

Secure ingestion

PDFs, DOCX, Markdown, CSV, JSON. Parsing, chunking with overlap, metadata extraction, vector embeddings with Amazon Titan. All inside your AWS VPC.

02

Retrieval with Row-Level Security

Semantic search over pgvector in Aurora PostgreSQL with multi-tenant isolation verified by automated tests. Metadata filters for your vertical or tenant.

03

Answers with Claude via Bedrock

Claude Haiku 4.5 / Sonnet 4.6 / Opus 4.6 with automatic cost-aware routing. Guardrails against cross-tenant leaks, prompt injection, and evidence-free responses. Real-time streaming.

04

Verifiable citations

Every answer links to the exact chunk in the source document. Forensic auditability reconstructible months later: who asked, which chunks were shown, which model answered, how much it cost.

05

Enterprise governance

Workspace RBAC, per-tenant policies, append-only audit log, per-query cost tracking, configurable retention, GDPR export, SAML/OIDC SSO, BYO KMS keys.

06

BYO-LLM and data residency

Plug in your own Anthropic, OpenAI, or private model credentials. Deploy in us-east-1 or eu-west-1. Zero vendor lock-in, zero transatlantic data transfer.

Who it's for

Built from day one for organizations that see their data as competitive advantage and legal obligation — not fuel for public SaaS.

✓

Multinationals with regulated data

Banking, insurance, healthcare, energy, corporate legal. Industries under GDPR, HIPAA-adjacent, SOX, PCI-DSS. They need AI over their knowledge without a single record leaving their AWS perimeter.

✓

Large enterprises with sensitive IP

Manufacturing, pharma, engineering, defense, strategy consulting. Patents, formulas, proprietary methods, strict NDAs. ChatGPT Enterprise is not legally viable.

✓

Chains and franchises with proprietary methodology

50+ locations needing operational consistency, standardized training, and answers aligned with HQ methodology at every point. AI distributes the corporate know-how.

✓

Mid-sized organizations scaling up

200–2000 employees with enterprise-equivalent compliance but without the Glean budget ($50k+ entry). The exact gap rags.cc fills.

✓

Public and semi-public sector

Municipalities, regulators, public universities, public hospitals. Non-negotiable data sovereignty, tender budgets, multi-year contracts.

✓

Teams already deep in AWS

Any organization that already chose AWS as its primary cloud and requires its AI tooling to respect its perimeter, IAM roles, KMS keys, and CloudTrail.

✕Who it's NOT for

  • Individual users chatting with personal notes (use NotebookLM)
  • Startups under 50 people with no regulatory requirements
  • Companies comfortable uploading data to ChatGPT Enterprise or Copilot
  • Teams not on AWS or with cloud-agnostic strategy
  • Cases where LLM creativity matters more than traceability
Deployment models

Three ways to keep your data protected

Pick the model that fits your risk profile, your current cloud, and your budget. All three honor the fundamental rule: your content never trains a third party and inference never goes to an unauthorized public SaaS.

Starter · Pro · Businessavailable

Shared multi-tenant

Operated by rags.cc in AWS us-east-1

The fastest way to get started. Sign up, pay by card, start uploading documents in minutes. We run the infra.

  • ✓Infra running in our dedicated rags.cc AWS account
  • ✓Multi-tenant isolation with Row-Level Security verified by automated tests
  • ✓Inference via Amazon Bedrock with private VPC endpoint (data never touches public internet)
  • ✓Encryption at rest with KMS managed by rags.cc
  • ✓Append-only audit log, exportable on request
  • ✓Anthropic does NOT use your data to train (Bedrock policy)
  • ✓rags.cc NEVER sells, transfers, or trains models with your content
  • ✓Optional BYO-LLM on Business (your own Anthropic / OpenAI credentials)
Ideal: Mid-sized companies, scaling organizations, and anyone that prioritizes time-to-value over owning the physical perimeter.
Enterprise+available

Dedicated deployment in your AWS account

Your account. Your VPC. Your KMS. Your CloudTrail.

We deploy the rags.cc stack literally inside your AWS account via Terraform and a limited cross-account IAM role. Your data never exists outside your perimeter.

  • ✓Infra inside the AWS account you authorize
  • ✓Your data physically lives in your Aurora, your S3, your VPC
  • ✓BYO KMS keys: the cryptographic root belongs to you, rotate whenever
  • ✓Your CloudTrail sees every operation; your security team audits everything
  • ✓rags.cc operates via temporary, auditable assume-role, with no direct data access
  • ✓You can revoke our access at any moment without losing your data
  • ✓SSO SAML/OIDC with your corporate IdP
  • ✓Negotiable SLA, 24/7 support, dedicated CSM
Ideal: Banking, insurance, healthcare, defense, public sector, regulated multinationals, contracts with mandatory external audit.
Business+ · EnterpriseQ4 2026

Federated connector (multi-cloud and on-prem)

For Azure, GCP, on-prem, and multi-cloud. Your documents are never copied.

We deploy a lightweight agent inside your infrastructure. Documents stay in your storage. On query, we fetch the required chunks on-demand through the agent via mTLS.

  • ✓Agent in Docker, systemd, or native binary; outbound-only, no inbound ports
  • ✓Connectors: SharePoint, Confluence, Azure Blob, GCS, S3, NFS, SMB, local folders
  • ✓Embeddings generated locally (bge-m3 on CPU) or via your cloud (Bedrock, Azure OpenAI)
  • ✓Documents are NEVER copied to rags.cc: only vectors and structural metadata
  • ✓At query time, selected chunks travel ephemerally to the LLM and are discarded
  • ✓mTLS with internal rags.cc CA + short-lived JWT; end-to-end encrypted traffic
  • ✓Auditable by your security team: local agent logs + reports in rags.cc
  • ✓Optional L3 mode: the agent invokes a local LLM and only the answer returns (extreme compliance)
Ideal: Organizations on Azure, GCP, on-prem, hybrid, or multi-cloud that cannot move their documents to AWS. Sectors with data sovereignty regulations (EU public sector, regional banks).
CriterionMulti-tenant AWSDedicated in your AWSFederated connector
Inference never touches public internet✓✓✓
Data NOT used to train third parties✓✓✓
Documents NEVER copied to rags.cc—✓✓
Works without an AWS account——✓
Supports Azure, GCP, on-prem——✓
Row-Level Security multi-tenant✓n/an/a
Encryption at rest with your KMS—✓✓
Revoke access in 1 click—✓✓
No inbound ports openedn/a✓✓
Minimum contractmonthlyannualannual
Time to start< 1 hour2-4 weeks1-3 weeks
Price range$99-$1,499 USD/mo$5k-$25k+ USD/mo$3k-$15k+ USD/mo
AvailabilitytodaytodayQ4 2026
RAG implementation options

There's more than one way to connect AI to your documents

Retrieval-Augmented Generation (RAG) lets an AI answer from your documents instead of its general training. It can be built in several ways, each with a different balance of cost, speed, accuracy, and data residency. We offer five, and we help you pick the right one.

Standard RAGCore platform

Classic: retrieve, then answer

One search, one answer. Simple, fast, and predictable.

How it works
For each question, we retrieve the most relevant passages from your documents using both meaning and exact keywords, rerank them by relevance, and the AI answers with citations.
Best for
Direct questions answered by a single document, such as internal policies, manuals, and FAQs.
Keep in mind
It can fall short when the answer is spread across several documents, and it answers even when the evidence is weak.
Where your documents live: Your AWS account (or ours, on the shared plan)
Gemini File SearchOn request

Managed by Google

Upload your files and Google handles the rest. Up and running in minutes.

How it works
Your documents are uploaded to a Google-managed store (Gemini File Search). Google chunks, indexes, and searches them, and a Gemini model answers with citations.
Best for
Fast proofs of concept, teams without their own infrastructure, and Google Cloud shops that are fine hosting documents there.
Keep in mind
Your documents leave AWS and are hosted by Google. You get less control over retrieval (no configurable reranking or keyword search), and it only works with Gemini models. We enable it only with your explicit approval.
Where your documents live: Google, with your organization's explicit approval
Agentic RAGOn request

Agentic: multi-step research

A team of AI agents that plans, searches, spots what's missing, and searches again.

How it works
The question is broken into parts. Several agents search in parallel, even across collections, and a reviewer checks whether the evidence is enough. If something is missing, they search again before writing the answer.
Best for
Complex questions that span several documents or departments, such as analysis, comparisons, and research.
Keep in mind
It makes several AI calls for every question, even easy ones. That makes it the slowest option (tens of seconds) and the most expensive per query.
Where your documents live: Your AWS account
RecommendedIn pilot

Adaptive Evidence Lean

Adaptive Evidence RAG · Lean

Our approach: effort scaled to difficulty, and no claim without evidence.

How it works
A classifier decides how much effort each question deserves. Simple questions with clear evidence are answered right away by a fast model. Complex ones trigger short research of up to two rounds. Before delivery, the riskiest claims are verified, and if information is missing, the answer says what's missing.
Best for
Most companies in production: high volume, mostly simple questions, and answers you can trust without runaway costs.
Keep in mind
It sometimes says "not enough evidence" instead of improvising. That's by design: we'd rather tell you what's missing than make something up.
Where your documents live: Your AWS account
Adaptive Evidence RAG · StrictIn pilot

Adaptive Evidence Strict

The same approach with maximum verification, for when a wrong answer is costly.

How it works
Like Lean, but our most accurate model writes every answer, and every claim is checked against its source before delivery. Anything unsupported is removed.
Best for
Healthcare, legal, compliance, insurance, and any collection flagged as sensitive.
Keep in mind
It costs more per query and takes a few seconds longer than Lean, in exchange for the lowest rate of wrong answers.
Where your documents live: Your AWS account

At a glance

CriterionClassicManagedAgenticAdaptive LeanAdaptive Strict
Cost per queryMediumVery lowHighMediumMedium-high
Response speedFastVery fastSlowFastModerate
Questions spanning several documentsLimitedLimitedVery goodVery goodVery good
Protection against made-up answersBasicBasicMediumHighHighest
Tells you what information is missing——Partial✓✓
Your documents stay in AWS✓—✓✓✓
Time to launchDaysMinutesWeeksDaysDays

Indicative comparison based on our internal cost and quality model. Before recommending an option, we validate it with a benchmark on your own documents and questions.

Every option includes the essentials: per-user and per-team permissions, citations to the original source, and an audit log for every answer.

Optional capabilities

Enabled only when your use case calls for it.

Jev · TypeSafe AIOn request · under evaluation

Fast decisions with Jev

Jev is a model built to make decisions, not to write. It can classify each question, decide whether the evidence is enough, and check every claim in the answer, while Claude keeps writing the response. Verifying every answer then costs a fraction of the price, and responses arrive sooner.

Applies to
Adaptive Lean
Keep in mind
Questions and cited passages are processed outside AWS, by TypeSafe AI. We only enable it with your explicit approval, never on sensitive collections, and only after validating it with questions in your language.

Not sure which one fits?

Tell us what your team asks and how sensitive your data is. We'll recommend an option and test it on your documents.

Talk to a specialist →
Under the hood

How RAG works, piece by piece, in plain English

Every answer is the result of several mechanisms working together. Here's what each one does, why it matters, and which options use it.

1. Preparing your documents

Before answering anything, your documents need to become something the AI can search.

Chunking

In plain terms: We split each document into manageable passages that follow its sections and headings.

Think of it as
Dividers in a filing cabinet: you search the right folder, not the whole cabinet.
Why it matters
An AI can only read so much text per question. Well-cut passages mean it gets exactly what's relevant.
Context on every passageContextual chunks

In plain terms: We add a short note to each passage saying where it comes from: which document, which section, and what it's about.

Think of it as
A caption on a loose photo that says "Ana's wedding, 2019". Without it, the photo loses its meaning.
Why it matters
A passage like "add 5 lb when…" is useless if you can't tell which protocol it belongs to. This step sharply reduces failed searches.
Meaning fingerprintsEmbeddings

In plain terms: We turn each passage into a list of numbers that captures its meaning rather than its exact words.

Think of it as
Map coordinates: texts about the same thing end up close together, even when they use different words.
Why it matters
It finds "how do I request vacation" even when the document says "time-off application".
Tags and filtersMetadata

In plain terms: Each passage carries tags such as client, department, date, language, or confidentiality level.

Think of it as
The filters in an online store: size, color, and price before you browse product by product.
Why it matters
It narrows the search to what's relevant and makes sure no one sees content they aren't allowed to see.
Managed storeGemini File Search store

In plain terms: Instead of us indexing the documents, they're uploaded to a Google store that handles chunking and search.

Think of it as
Renting a warehouse with inventory service included: convenient, but your goods sit in someone else's warehouse.
Why it matters
It removes infrastructure and speeds up launch, at the cost of control and of hosting documents outside AWS.
Used in
Managed

2. Finding the right information

When a question arrives, we need the passages that actually answer it.

Semantic searchVector search

In plain terms: We look for the passages whose meaning fingerprint is closest to the question.

Think of it as
Asking a librarian about a topic instead of looking up a word in the index.
Why it matters
It understands synonyms and paraphrasing, but it can miss proper names, codes, or acronyms.
Keyword searchBM25

In plain terms: We look for the passages that literally contain the question's key terms.

Think of it as
Good old Ctrl+F, but smart enough to know which words matter most.
Why it matters
Essential for author names, legal article numbers, product references, and technical jargon.
Hybrid searchRRF fusion

In plain terms: We combine both searches and keep what both rank highly.

Think of it as
Asking two experts with different criteria and going with what they both recommend.
Why it matters
You get the best of both: it understands what you mean without losing exact terms.
Reranking

In plain terms: A specialized model rereads the top candidates alongside the question and sorts them by true relevance.

Think of it as
A second interview round: the first screen picks 20 candidates, the second picks the best 5.
Why it matters
The AI gets less noise and more precision, and the score tells us how solid the evidence is.

3. Thinking before answering

Not every question deserves the same effort. These mechanisms decide how much research to do.

Question classifierRouter

In plain terms: Before answering, we classify the question as simple, compound, or multi-step. Search starts in parallel so no time is lost.

Think of it as
Triage in an emergency room: not every patient needs the same specialist.
Why it matters
It avoids overspending on easy questions and underdelivering on hard ones.
Evidence gate

In plain terms: We check whether what we found is enough to answer. If it is, we answer; if not, we dig deeper or say so.

Think of it as
An editor who won't run a story without at least one solid source.
Why it matters
It's the main defense against made-up answers, and it adds almost no cost because it reuses signals we already have.
Multi-step researchAgentic loop

In plain terms: We split the question into parts, search for each one, check what's missing, and repeat if needed.

Think of it as
A researcher who reads a report, realizes another fact is needed, and goes to find it.
Why it matters
It handles questions whose answer is spread across several documents. In Adaptive we cap it at two rounds to keep time and cost in check.
The right model for each questionModel routing

In plain terms: Simple questions with clear evidence are written by a fast model; complex ones by the most accurate one.

Think of it as
You don't need a surgeon to put on a bandage, but you do for an operation.
Why it matters
With good evidence, writing the answer is easy. Using the large model only when it adds value lowers cost and wait time.
Fast decisions with JevDecision model · optional

In plain terms: A model that doesn't write: it only answers closed questions, such as "is this simple or complex?" or "does this source support this sentence?", with a probability.

Think of it as
A line judge: they don't play the match, but they call in a split second whether the ball was in.
Why it matters
It makes verification cheap enough to check every claim. It's optional and only enabled with your approval, because it processes data outside AWS.

4. Answering with guarantees

An answer only goes out when every claim is backed up.

Cited answersCitations

In plain terms: Every claim links to the exact passage of the document it comes from.

Think of it as
The footnotes in a good report.
Why it matters
You can check any fact in one click and audit the answer months later.
Claim verificationEvidence Contract

In plain terms: Before delivering an answer, we check that each citation really supports what's being claimed. Anything unsupported is removed.

Think of it as
A fact-checker reviewing an article before it goes to print.
Why it matters
Other approaches check whether they found enough information; we also check the answer itself. Lean applies it to the riskiest answers, Strict to all of them.
Saying what's missingAbstention with gaps

In plain terms: When there isn't enough evidence, the answer says so and explains what information your knowledge base is missing.

Think of it as
A good advisor who'd rather say "I don't know, this piece is missing" than guess.
Why it matters
A made-up answer can be costly. A "this is missing" can be fixed: every gap tells you which document to add.
How we operate

Your sensitive data never leaves your infrastructure

Pick the case that matches your reality. In all three, the only thing rags.cc ever sees of your content is encrypted mathematical conversions — never your readable documents.

You keep the readable part

Your original documents — PDFs, contracts, emails, knowledge bases — stay where they live today. rags.cc does not copy them, does not relocate them, does not store them in its own account.

We only see vectors

So AI can search, we convert each fragment into a mathematical vector (1024 numbers). That vector is irreversible: no public algorithm reconstructs the original text from it. It's all rags.cc needs to find the right answer.

End-to-end encryption

Everything moves over TLS 1.3 with mutual certificates (mTLS) or private VPC. At rest, everything is encrypted with KMS — whenever possible, with your own keys (BYOK). Neither rags.cc nor AWS can decrypt without your cryptographic consent.

Auditable by your team

CloudTrail, connector logs, and rags.cc audit log stay in your infra. Your security team sees every action. You can revoke access in one click and keep all your data.

Client lives in AWS

You grant us controlled access to your account. We never see your readable documents.

Ideal: Companies already on AWS looking for fast time-to-start with their own perimeter.

  1. 1

    1. You authorize

    You create a cross-account IAM role in YOUR AWS account with minimal permissions (read from your designated S3, write to your designated Aurora). You can revoke it at any time.

  2. 2

    2. Deploy into your account

    rags.cc Terraform stands up Aurora + pgvector + ECS services INSIDE your account. Documents, chunks and embeddings physically live in your infra. rags.cc only operates via short-lived roles audited by your CloudTrail.

  3. 3

    3. Mathematical conversion

    When a document is uploaded, the worker parses and generates embeddings by calling Amazon Bedrock over a private VPC endpoint. Original text never leaves your VPC. Vectors are stored next to the document.

  4. 4

    4. Search and answer

    When a user asks, we convert the question into a vector and retrieve the most similar chunks FROM YOUR Aurora. Those chunks are sent to Claude (Bedrock) with a citation prompt. The answer comes back with every citation linked to the original document in YOUR S3.

  5. 5

    5. What rags.cc sees

    Nothing readable. Aggregated metrics (latency, cost, errors), technical logs, and temporary operational access via IAM. Zero content.

What exactly is an embedding?

An embedding is a 1024-real-number vector that represents the meaning of a text fragment. Two fragments with similar meaning produce vectors close together in mathematical space. It is an irreversible function: given the vector, no public algorithm recovers the original text. It is all rags.cc needs to find, but not to read.

rags.cc stores and searches over these vectors. Never over the text.

Example

📄 "The Globex lease contract expires on December 15, 2026"

↓ embed v2

[0.0213, -0.4571, 0.1084, 0.9372, …] (1024 numbers)

Does your case fit one of these three?

Tell us your concrete requirement and we spin up a demo environment for your company. We're operating invite-only until we complete our first cohort.

Request early access →
Architecture

Serious architecture. Documented decisions.

Every technical choice we make is written down, justified, and reviewable. That's how we turn "your data is safe" into something you can verify — not a marketing promise.

Your data never leaves your network

rags.cc runs inside your AWS VPC. Calls to AI models travel through private AWS endpoints — your bytes never touch the public internet. Not even we can inspect them.

Isolation between customers — verified, not assumed

Row-Level Security in PostgreSQL: even a bug in our code cannot read another customer's data. The database blocks the leak before the application even tries the query.

Every answer links to the exact paragraph

Not a vague summary. Every claim our agent makes cites the document chunk it came from, with a stable ID. You can audit an answer six months later and reconstruct exactly what the model saw.

We don't lock you to a single AI provider

We start with Claude via AWS Bedrock because it honors our data residency promise. If you want Llama, Mistral, or your own instance tomorrow — our Model Gateway supports it without rewriting your integration.

Built for your auditors from day one

Separate encryption per data type. Automatic credential rotation. Immutable audit logs. Backups encrypted with your own key if needed. Meets the principles of SOC2 and ISO27001 — not as an afterthought, but in every line of Terraform.

Start cheap. Scale linear. No lock-in.

Our architecture scales elastically: Aurora Serverless, pay-per-use inference, auto-scaling storage. You can start with a few documents and grow to millions without re-architecting.

Want the full technical brief?

We publish every Architecture Decision Record. Ask us for the 30-min walkthrough where your team asks anything. No NDAs.

Book a walkthrough →

Why rags.cc

Private by design

Everything runs in your AWS VPC. Bedrock via private endpoint. Zero egress.

Verifiable citations

Every answer links to the exact chunk. Forensic auditability months later.

True multi-tenant

Row-Level Security in PostgreSQL. Isolation verified by automated tests.

LLM replaceability

Decoupled Model Gateway. Switch from Claude to Llama without rewriting the product.

Simple public pricing

All prices in USD. No surprises. Monthly or annual billing with a discount.

Starter

$99 USD/ mo
  • 1 workspace
  • 500 documents
  • 1,000 queries / mo
  • 10M LLM tokens / mo
  • Claude Haiku 4.5
  • 30-day audit log
Request
★ Popular

Pro

$499 USD/ mo
  • 5 workspaces
  • 5,000 documents
  • 10,000 queries / mo
  • 50M LLM tokens / mo
  • Haiku 4.5 + Sonnet 4.6
  • 90-day audit log
  • 99.5% SLA
Request

Business

$1,499 USD/ mo
  • 25 workspaces
  • 50,000 documents
  • 50,000 queries / mo
  • 200M LLM tokens / mo
  • Haiku + Sonnet + Opus 4.6
  • BYO-LLM
  • 1-year audit log
  • 99.9% SLA
Request

Prices shown in US dollars (USD). Taxes not included. Monthly or annual billing with a discount. Enterprise tier with SSO, dedicated deployment, and 99.95% SLA available under contract.

Talk to us →