What is a vector database?

A database that stores meaning and finds what is close to it.

An embedding model turns a sentence, a product photo or a support ticket into a list of numbers — a vector — where similar things end up near each other. A vector database stores millions of these vectors and answers one question very fast: which stored items are closest to this one? That is what lets a chatbot find the right policy paragraph, or a shop show "more like this", even when the words do not match.

For large language models it is the memory layer behind retrieval-augmented generation (RAG): retrieve the most relevant chunks of your own documents, hand them to the model, and get answers that cite your sources. XcellAI Vector Database gives Indian teams in retail, healthcare, BFSI and SaaS this layer as a managed service — sized, indexed, secured and run in India by XcellHost.

Supports leading PostgreSQL + pgvector* Open-source vector engines* OpenSearch k-NN* LangChain & LlamaIndex* Open & hosted embedding models*
01Built for AI workloads

Stores embeddings next to their text and metadata, so retrieval for search and RAG is one call, not three systems.

02Memory for LLMs & chatbots

Feeds the right passages into the prompt, so answers come from your documents and can quote them.

03Grows with your corpus

From a single node for a pilot to sharded, replicated clusters when the collection reaches hundreds of millions of vectors.

04Ready for production

High availability, role-based access, encryption, backups and monitoring set up before go-live — not after an incident.

05Real-time retrieval

Approximate nearest-neighbour indexes return top matches in milliseconds* while new vectors keep arriving.

06Fits your AI stack

REST, gRPC and Python clients, plus integrations with common RAG frameworks and the models you already use.

Not sure which engine or how much RAM?Our sizing calculator and engine picker answer both in under two minutes.
Size it now →
How a vector database works

Six steps from raw content to a ranked answer.

Pick a step to see what happens there. The diagram on the right follows along.

embed [0.12, −0.48 …][−0.07, 0.33 …][0.64, 0.02 …] In memory On NVMe disk id · vector · metadata layer 2layer 1layer 0 query vector → top-k by cosine similarity vector BM25 filterregion = IN · year ≥ 2025 1 · Leave policy 2026 §4 2 · HR handbook p.12 3 · Circular 07/2025 vector DBtop 5 chunks RAG chatbot semantic search recommendations
Turn content into vectors

—

Key benefits

What a managed vector database changes for your team.

Fast similarity search

Approximate nearest-neighbour indexes skip most of the collection, so top-k results return in milliseconds* instead of scanning every row.

recall vs latency, tuned

Made for ML and LLMs

Embeddings from text, image and audio models are first-class data, stored with the chunk text and source link your app needs.

any embedding model

Scales without a rewrite

Start on one node; add replicas for read load and shards for size. Your application keeps the same endpoint.

replicas · shards

Better context for answers

Search by meaning plus keyword and filters means the model sees the right paragraph — fewer made-up answers.

hybrid + re-rank

Security and compliance built in

Private networking, encryption in transit and at rest, role-based access and audit logs, with data kept in India.

DPDP-ready design

Plugs into your AI ecosystem

Works with common RAG frameworks, notebook tools and the model endpoints on our AI platform or the public clouds.

REST · gRPC · Python
Better retrieval is the cheapest way to better AI answers.Start with a sizing review — we come back with an engine, index and INR quotation.
Why it matters · indexes & search

Ordinary databases match exact values. AI needs nearest meaning.

Why a vector database is important

Embeddings have hundreds to thousands of dimensions. Comparing a query against every one of millions of rows is too slow for a chatbot, and a B-tree index cannot help. A vector database keeps a purpose-built index for this, plus the filters, access control and backups an enterprise system needs.

Deploying one, without the guesswork

The hard parts are choices: which engine, which index, how much RAM, how to chunk documents and how to mix keyword with vector search. XcellHost makes those choices with you, deploys on infrastructure in India and runs it — so your developers call one endpoint from the framework they already use.

Hybrid search, livesample query · "leave encashment rules"
BM25 keyword
exact terms
Vector
meaning
Filter
dept = HR
Fused ranking · reciprocal rank fusion, k = 60
  1. Leave policy 2026 · §4 Encashment0.0325
  2. Leave FAQ · carry-forward limit0.0323
  3. HR handbook · p.120.0318

HNSW · graph index

A layered graph: search enters at a sparse top layer and walks closer neighbours layer by layer. Very good recall at low latency; the graph lives in RAM, so memory is the main cost.

  • Best default for most RAG collections
  • Tuned with M, ef_construction and ef_search
  • Plan roughly 1.3–2× the raw vector size in RAM*

IVF and product quantisation

Vectors are grouped into clusters; a query only checks the closest few clusters (nprobe). Product quantisation compresses each vector into short codes, cutting memory many times at some cost in recall.

  • Lists and nprobe set the speed/recall trade-off
  • PQ or int8 compression for large collections
  • Re-train when the data distribution shifts

DiskANN · SSD-resident graph

The full graph and vectors sit on fast NVMe, with only compressed codes in RAM. Suits hundreds of millions to billions of vectors where an all-in-RAM index would be too costly.

  • Most data on NVMe, small RAM footprint
  • Billion-scale collections at lower cost
  • Latency depends on SSD performance

Hybrid · BM25 + vectors

Keyword search catches exact names, codes and numbers that embeddings blur; vector search catches paraphrases. Running both and fusing the ranks usually beats either alone for enterprise RAG.

  • BM25 or sparse vectors for exact terms
  • Reciprocal rank fusion or weighted blend
  • Optional cross-encoder re-ranker on the top 50

Metadata filters & multi-tenancy

Store department, language, date, customer or access label with each vector and filter during the search, not after it. For SaaS products, each customer gets a namespace, partition or collection.

  • Pre-filtering keeps top-k complete
  • Per-tenant namespaces or collections
  • Document-level access labels enforced at query time
Use cases

Where Indian teams use vectors first.

Hover or tap a card to see what is stored and what comes back.

RAG chatbots on company knowledge

Policies, SOPs and manuals answered with citations for staff or customers.

Storesdocument chunks + source, page, access labelReturnstop 5 passages for the LLM prompt
E-commerce search & "more like this"

Shoppers type what they mean; similar products appear even with different words.

Storesproduct text and image embeddings, price, stockReturnssimilar, in-stock items under a price filter
Healthcare knowledge retrieval

Clinicians and staff find protocols and prior notes by meaning, with data kept in India.

Storesde-identified notes, guidelines, formularyReturnsrelevant protocol passages with sources
BFSI document intelligence

Circulars, KYC and loan files searched by meaning; similar-case lookup for analysts.

Storescirculars, policy clauses, case summariesReturnsmatching clauses filtered by date and regulator
Support deflection & agent assist

New tickets matched to solved ones; agents get suggested answers as they type.

Storesresolved tickets, KB articles, product tagsReturnssimilar solved tickets and the fix used
Duplicate & anomaly detection

Near-duplicate invoices, listings or claims flagged because their vectors sit too close.

Storesinvoice, listing or claim embeddingsReturnspairs above a similarity threshold
How XcellHost delivers

The engine is open source. Running it well is the service.

Eight things you get from XcellHost that a self-installed vector database does not come with.

Tuned searchIndex parameters set against your own recall and latency targets.
AI-ready storageNVMe for indexes, S3-compatible object storage for backups and raw files.
Room to growReplicas and shards added as the collection and query load grow.
Semantic qualityChunking, embedding model and hybrid weights reviewed with your test questions.
Cloud-nativeContainers on managed Kubernetes or VMs, with a private endpoint for your apps.
Security & complianceRBAC, TLS, encryption at rest, audit logs, watched by our 24×7 SOC.
Flexible APIsREST, gRPC and Python clients; SQL too when the engine is pgvector.
Cost in rupeesRight-sized RAM and compression; one INR GST invoice.
Engine & index chosen with you

pgvector, a dedicated vector engine or OpenSearch — picked on scale, features and skills.

01
Embedding pipelines

Ingest from file shares, SharePoint, S3 buckets or databases; chunk, embed and upsert on a schedule.

02
24×7 NOC + SOC, Mumbai

Monitoring, patching, capacity alerts and incident response by our own engineers.

03
Backups & restore drills

Snapshots to S3-compatible storage in India, with restores tested on a schedule.

04
GPU when you need it

Embedding and re-ranking on NVIDIA L4, RTX PRO 6000 or H200 servers — priced on quote.

05
Monthly report

Vectors, query volume, p95 latency, recall checks and capacity forecast in plain language.

06
India compliance

Embeddings are derived from your data. Treat them like your data.

Vectors can leak meaning about the text they came from, so they deserve the same care as the source documents. Hosting and operating them in India helps you meet the rules below — compliance itself stays with your organisation.

DPDPDPDP Act 2023 & Rules 2025

Personal data inside chunks and metadata is processed under your instructions, with access control, deletion by document ID and breach support. As the Rules phase in, the Board must be intimated of a breach without delay, with a detailed report within 72 hours.

6 hCERT-In directions (April 2022)

Reportable cyber incidents must be reported within 6 hours of noticing them, and ICT system logs kept for a rolling 180 days in India. Query and access logs can be retained in our Indian data centres.

INData residency

Vectors, source text, backups and logs held in XcellHost's Mumbai Tier IV data centre and/or partner DCs in India*, on infrastructure our own teams run.

BFSISector expectations

RBI and SEBI expect outsourced IT to be governed, audited and exit-ready. We provide access logs, change records and a documented hand-over path.

Regulatory summaries checked October 2026 for orientation only — not legal advice. Confirm current requirements with your counsel, CERT-In, the Data Protection Board, RBI or SEBI.

Plans

Priced by index memory, per month. Managed end to end.

Index RAM is the main driver of cost, so plans are sized by it. Embedding model usage and GPU servers are separate.

Starter

Pilot & small RAG

₹9,999/month*
single node · up to 8 GB index RAM · 100 GB NVMe
  • Managed pgvector or a dedicated vector engine*
  • Vector + keyword search, metadata filters
  • Daily backups, 7-day retention
  • TLS, encryption at rest, API keys
  • 24×7 monitoring, business-hours support
Most chosen Professional

Production RAG & search

₹39,999/month*
3-node HA · up to 64 GB index RAM · 500 GB NVMe
  • Everything in Starter
  • High availability with automatic failover
  • Hybrid search and re-ranking set up
  • RBAC, multi-tenant namespaces, private endpoint
  • Hourly snapshots, 30-day retention
  • One managed embedding pipeline
  • Monthly report, 24×7 support
Enterprise

Large & regulated

On quoteafter a sizing review
sharded clusters · 100M+ vectors · DiskANN
  • Everything in Professional
  • Sharding, DiskANN and quantisation
  • GPU embedding / indexing — price on quote
  • DR copy in a partner DC in India*
  • Named engineer, custom SLAs
  • Audit evidence packs for BFSI

* Indicative — confirm with XcellHost. INR, excluding 18% GST. Backup storage is ₹1,999/TB/month and extra embedding pipelines ₹2,999/month, the same rates as our Managed Data Lake page; setup ₹24,999 one-time matches our operations onboarding fee. Engine availability per plan is confirmed in your quotation. 99.95% uptime SLA applies to HA plans.

Configure

Build your vector database plan. See the rupee total.

Fixed-price items go to the cart. Enterprise and anything with GPUs comes as a quotation.

1 · Plan
5 · One-time services
Your estimate · indicative

Professional · 64 GB RAM

    Monthly₹0
    One-time₹0
    Subtotal₹0
    GST (18%)₹0
    First invoice incl. GST₹0
    * Indicative — confirm with XcellHost. INR excl. 18% GST. GPU configurations are priced on quote and never added to the cart.
    Tool · vector storage & RAM sizing

    How much memory does your index need?

    A rule-of-thumb estimate. Every input, including the index overhead factor, is yours to change; the formula sits under the result.

    Dimensions
    Precision (bytes per dimension)
    Index type (sets a default overhead)
    Estimated footprint
    RAM, all replicas0 GB
    Disk, all replicas0 GB
    Index in RAM 0Metadata (RAM) 0Vectors on disk 0
    Fits Professional
    Formula vector data (GB) = vectors × dimensions × bytes ÷ 10⁹ index in RAM = vector data × overhead factor metadata = vectors × metadata bytes ÷ 10⁹ RAM = (index in RAM + metadata) × replicas disk = (vector data + index in RAM + metadata) × replicas —
    Illustrative only. Default overheads (HNSW 1.5×, IVF 1.1×, DiskANN 0.25× — only compressed codes in RAM) are rules of thumb; real figures depend on engine, M / nlist settings and quantisation. We size precisely during the review.
    Tool · engine picker

    Five questions. Which engine fits?

    We run leading open-source engines*. Answer for your project and the ranking updates as you go.

    Best fit for your answers

      —

      Suggested plan · Professional
      * Named projects are open-source software used to describe compatibility; trademarks belong to their owners and no endorsement is implied. Scores are a guide, not a benchmark.
      Compare

      Self-hosted, a foreign SaaS, or XcellHost managed.

      What you needInstall it yourselfOverseas vector SaaSXcellAI Vector Database
      Where vectors liveWherever you set it upOften outside IndiaMumbai Tier IV / partner DCs in India*
      Choosing engine and indexYour team researches itOne engine, fixedPicked with you from leading engines
      Sizing RAM and diskTrial and errorPay for the tier aboveSized from your data, reviewed monthly
      3 a.m. outageWhoever is on callTicket in another time zone24×7 NOC in Mumbai
      Backups and restoreScripts, rarely testedVaries by planSnapshots in India, restore drills
      Embedding pipelinesBuild and babysitBring your ownBuilt and scheduled by us
      BillingHardware + peopleUSD, card, forex swingsOne INR GST invoice
      Get started

      Live in about three weeks. Tested on your questions.

      Typical timings for a Professional deployment up to a few crore vectors*.

      1. 01ReviewData sources, volumes, query patterns and 30–50 test questions agreed.Days 1–3*
      2. 02DesignEngine, index, chunking, embedding model and tenancy model chosen.Days 3–6*
      3. 03DeployCluster, private endpoint, RBAC, encryption and backups in India.Week 2*
      4. 04Load & indexPipelines ingest, embed and upsert; indexes built and tuned.Week 2–3*
      5. 05EvaluateRecall and latency measured on the test questions; hybrid weights set.Week 3*
      6. 06Go live & runHanded to the NOC; monthly report and capacity forecast.ongoing