A language model on its own answers from what it absorbed in training — it has never read your credit policy, your SOPs or last week's board circular. Retrieval-augmented generation (RAG) fixes that at answer time: the system finds the most relevant passages in your approved content and gives them to the model with the question, so the reply is built from your material and cites it. In an enterprise, that retrieval also has to honour who is asking.
XcellAI Enterprise RAG as a Service runs on Sarv's Knowledge & RAG platform — knowledge bases, hybrid retrieval, retrieval testing and agent connectors. XcellHost adds the enterprise layer around it: identity and permissions, connectors, governance, evaluation, private hosting in India and a 99.95% SLA. Per Sarv, the platform can pair retrieval with fine-tuning for domain assistants, as in its Kar Saathi deployment.
“What is the approval limit for a branch manager?”
“Approval limits vary by bank, typically…”
“What is the approval limit for a branch manager?”
“Up to ₹25 lakh for secured retail loans [1]; above that, the regional credit committee approves [2].”
Example is illustrative. * XcellHost delivery or integration scope, confirmed per engagement — not a published Sarv capability. Sarv lists websites, documents and Markdown as source types.
Sarv's knowledge-base controls cover ingestion, chunk curation, retrieval testing and agent connectors. We walk through each with your owners, security and compliance teams before anything goes live.
Choose the websites, documents or Markdown that are approved to answer from. Check extracted text, headings and tables before they become chunks.
Try hybrid, text-only and vector search side by side. Compare which chunks come back, their rank, relevance and whether they are active.
Edit weak chunks, write FAQs for real questions, switch on only the source subset you trust, and keep an owner and version review for each.
Confirm embedding coverage and which embedding model was used — separately from "job finished". Test response time and cost at production-like volumes.
Write down the evidence each common, ambiguous and unanswerable question should use. Judge retrieval first, then the wording of the answer.
Decide which users and agents may read which source, where data is hosted, and how edits and deletions in a source reach the index.
One "looks good" demo hides three different failure points. Each gets its own acceptance check: chunk review, retrieval testing, the agent connector, and change monitoring on the specific pages or whole site you choose.
The quickest way to know whether RAG fits is to try it on your own content, questions and access rules — not a vendor demo.
A representative slice: policies, manuals, an FAQ site, a scanned PDF — with the access groups that apply to each.
Questions people actually ask — from tickets, call notes or the team inbox — including ones that should be refused.
Who asks, where (intranet, Teams, web, WhatsApp, phone), and what happens when the assistant cannot answer.
Tap what matters for your project — the recommendation moves.
Supplies current information from approved sources with each question. Update or restrict a document and the next answer changes — no retraining.
Fine-tune for the vocabulary and style of a field, retrieve for the facts that change. Per Sarv, its Kar Saathi deployment combines the two.
Trains a model on examples so it follows a format, tone or task. It does not keep facts current, cannot enforce per-user access, and cannot show where an answer came from.
How retrieval supports an AI answer, in the order we build it — with the enterprise checks added at each step.
Pick approved content, import or connect it, inspect the chunks, and keep each chunk's source, revision and access label visible so every answer can be traced.
Run real questions in the retrieval bench, compare hybrid, text and vector modes, narrow by knowledge base or URL filter, and test as users in different groups.
Attach the knowledge-base connector to an AI agent, add guardrails, then score answers against the golden set and keep scoring after launch.
From a shared service for a first assistant to hardware that never leaves your building. Slide across to compare.
Private hosting runs on XcellHost Private Cloud; GPU capacity for open models comes from AI Factory — see the NVIDIA GPU comparison. GPU is priced on quote. * Options B and D, and component placement for Sarv's platform, are confirmed per engagement.
If a person cannot open a document in the source system, the assistant should not quote it to them. We design for that from day one.
SAML 2.0 or OpenID Connect against your identity provider, including Microsoft Entra ID — run with our Managed Identity Service. MFA and conditional access stay where they are.
Admins, knowledge owners, reviewers and end users get separate rights. Knowledge bases are scoped to groups, and every admin change is logged.
Access labels come from the source system's ACLs at sync time and are applied as a filter during retrieval, so restricted chunks never reach the prompt.
“What changed in the bonus policy this year?”
Connectors for leading enterprise systems* bring content, metadata and permissions into the index on a schedule, so nobody uploads files by hand.
* Connector availability, sync method and permission mapping are confirmed per system during discovery; some connectors are built and run by XcellHost. Product names belong to their owners. Enterprise plan includes 5 connectors; more at ₹9,999* each per month.
Each control is agreed in writing, tested in the pilot and evidenced in the monthly report.
Who asked what, which sources were retrieved, which model answered and what was shown — exportable to your SIEM.
Personal identifiers masked in prompts and logs before the model sees them. Pair with Scrutiny DLP across endpoints and email.
Separate retention for sources, chunks, prompts and logs — for example, 180 days for logs in India.
Topic policies, prompt-injection checks, refusal for out-of-scope questions and a hand-off to a person.
Hosted models or private open models on GPU in India — chosen per use case. See OpenAI and Gemini managed services.
Your content and prompts are not used to train models; we confirm the terms of every model provider in the contract.
TLS in transit and encryption at rest, with keys you control through Cloud HSM (BYOK). Revoke the key and the data becomes unreadable.
* XcellHost delivery scope, confirmed per engagement and deployment model — not a published Sarv capability.
We measure before launch and keep measuring, so a new source, model or prompt never quietly makes answers worse.
Filter by team. Each card shows what is indexed and the control that matters most.
Credit, operations and branch staff ask about policies and circulars; answers cite clause and version.
Leave, payroll, device and access questions answered from handbooks and the IT knowledge base, inside Teams or the intranet.
Find indemnity caps, renewal terms and obligations across contract libraries, with the exact clause quoted.
Agents get a drafted, cited answer as the ticket or call arrives — on web, WhatsApp or voice via DeepCall AI.
Runbooks, design docs, incident reviews and tickets searchable in plain language for engineers and NOC teams.
Compliance teams map obligations to internal controls and ask "what does this circular require of us?"
Sarv provides the Knowledge & RAG platform. XcellHost scopes, implements, secures, bills and supports it — with an SLA you can hold us to.
Use cases, sources, users, data classes and a measurable target agreed in writing.
01SSO, groups and source permissions mapped and tested as real users.
02A golden set scored before launch and after every change.
03Intranet, web, WhatsApp or voice through DeepCall AI.
04We take the first call and escalate to Sarv when the platform needs it.
05Platform, delivery and support on one GST invoice; model usage shown separately.
06Chunks, embeddings, prompts and logs can all carry personal data. The way we set up enterprise RAG helps you meet the rules below — compliance itself stays with your organisation.
Index only data with a lawful purpose, honour erasure by removing and re-indexing, and keep access per source. As the Rules phase in, a breach is intimated to the Board without delay, with a detailed report within 72 hours.
Reportable incidents go to CERT-In within 6 hours of noticing them, and ICT logs are kept for a rolling 180 days within India. Audit and access logs are planned for that.
Outsourced technology must be governed, auditable and exit-ready — RBI's IT outsourcing directions, SEBI's CSCRF and IRDAI's guidelines. You get source lists, access records, test results and an exit plan.
The AI management system standard. Enterprise RAG evidence — risk register, evaluation, monitoring — feeds your programme; our ISO 42001 consulting helps you build it.
Regulatory summaries checked October 2026 for orientation only — not legal advice. Confirm current requirements with your counsel, CERT-In, the Data Protection Board or your sector regulator. * Placement of Sarv platform components and external model use are confirmed per engagement.
Each plan includes the Sarv platform, XcellHost delivery and support. LLM and embedding usage is billed at actuals or on your own keys; GPU for private models is on quote.
* Indicative — confirm with XcellHost. INR, excluding 18% GST. Sarv publishes no prices for RAG as a Service, so all figures are XcellHost placeholders. Business reuses the Professional price on our RAG as a Service page; add-ons reuse that page's prices. LLM and embedding usage, WhatsApp conversation charges and GPU are billed separately.
Fixed-price items go to the cart. Dedicated, private, voice and GPU come as a quotation.
Every assumption is editable. Model prices are yours to enter — we never assume them — and the working sits under the result.
pages = documents × pages/doc · tokens = pages × tokens/page
chunks = ⌈tokens ÷ (chunk × (1 − overlap))⌉
index = chunks × dims × bytes × overhead · text = chunks × chunk × 4 B
LLM tokens/month = queries × (k × chunk + prompt + answer)
cost = tokens ÷ 10⁶ × your ₹ rate
—
Pick what describes your organisation. The recommendation, reasons and plan update as you go.
| What you need | In-house RAG build | Public AI assistant | XcellAI Enterprise RAG |
|---|---|---|---|
| Who sees which document | Your team writes ACL filters | Whatever the user pastes in | Source permissions inherited at retrieval* |
| Where data is processed | Wherever you host it | Provider's regions, often abroad | Mumbai Tier IV or your VPC* |
| Citations & audit trail | Only if you build them | Rarely tied to your sources | Cited answers, every question logged |
| Personal data in prompts | Custom redaction code | Policy and training risk | PII redaction before the model* |
| Proof it works | Ad-hoc testing | None for your content | Golden set scored monthly |
| Uptime & support | Your on-call team | Consumer terms | 99.95% SLA, 24×7 NOC + SOC |
| Billing | Several USD vendors + people | Per-seat USD subscriptions | One INR GST invoice |
Typical timings for an Enterprise deployment on shared infrastructure*. Private deployments add hardware lead time.