It guesses
When retrieval finds nothing useful, a generative model fills the gap fluently and plausibly.
Deployed in production
AT
Ask Tara
An enterprise assistant that answers from your own knowledge — with a citation on every response, deterministic calculations it never invents, and a refusal when there is no reliable source. Built for organisations where a wrong answer is a compliance event, not a bad review.
9Languages — English plus eight Indian languages
100%Answers carry a citation, or the assistant declines
2Person approval on every regulated change
0Static secrets — managed identity throughout
The problem
Most enterprise chatbots are confidently wrong
A general assistant pointed at a document library will answer everything, including the questions it should refuse. In a regulated business that is not a quality issue — it is a customer being told the wrong loan-to-value ratio, an employee acting on a superseded policy, and no record of why.
It calculates
Ask a model for an eligibility figure and it will produce one. Sometimes it is even right.
It bleeds across domains
A lending question surfaces HR content because one index holds everything and nothing is scoped.
It cannot be corrected
Fixing a wrong prompt or a stale parameter means a software release and a change window.
Ask Tara is built the other way round. Grounding, refusal, deterministic computation and human approval are architectural constraints, not prompt instructions — because a prompt is a request and a constraint is a guarantee.
Capabilities
What Ask Tara does
Grounded answers with citations
Every response carries its source. If there is no reliable source, the assistant declines rather than guess. Answers stream in real time over SSE, so the experience is conversational without the model being unconstrained.
FAQ fast-path — approved wording, unaltered
A confident FAQ match is returned exactly as approved, with no AI rewording. Fast, precise and compliant — the answer legal signed off is the answer the user sees. Semantic matching means it works however the question is phrased, including typos and paraphrases.
Deterministic computation
Eligibility, loan-to-value tiers, slab limits, top-up caps and deviation buffers are computed by a rules engine driven by administrator-editable parameters — never by the model. The AI presents the result and cites it; it does not calculate it.
Multilingual by architecture
English plus eight Indian languages through a translation gateway — the knowledge base, FAQs and prompts stay English-only, so there is no dual-corpus drift. Compliance-critical phrases (refusals, consent, disclosures) come from a curated lexicon and are never machine-translated.
Voice, for people who do not type
Speech-to-text and text-to-speech, administrator-controlled. The hard part is not the engine — it is deterministic speakable text: currency symbols, operators, percentages and code-switched terms rendered so a neural voice reads them correctly, and citation URLs suppressed rather than read aloud.
Connector platform & governed actions
Protocol drivers are built once (REST and MCP); each integration is administrator configuration — endpoint, auth, field map, action items, department scope, routing order, confidence floor and budget. Write actions surface a plain-language confirm card with a tamper-evident token binding the exact arguments.
Automated knowledge ingestion
A nightly SharePoint sync that processes only changes, handles deletions and renames, and respects each document's audience and permissions — so entitlement is enforced at retrieval, not hoped for in a prompt.
Administration console
Role-based control over FAQs, the indexer, safety shields, languages, voice and configuration — plus "what people asked", which shows the question, the answer given and the outcome, with personal data redacted.
The answer model
Layered, not one shot
A question does not go straight to a language model. It descends a ladder, and stops at the first tier that can answer it correctly.
Safety checkscontent shields · prompt-injection defence · PII detection
↓
Understand the question and its parametersintent, domain, and the values a computation would need
↓
FAQ tier(a) exact approved answer · (b) reasoned over FAQ knowledge · (c) computed scenario
↓
Knowledge tierdomain-scoped, grounded, reasoned answer with a citation
↓
Otherwiseask for the missing detail · route to a person · politely decline
The last row is the one that matters. An assistant that cannot say "I don't have a reliable source for that" is not safe to deploy in a regulated business.
Knowledge governance
A front door for content, with no back door
Retrieval quality is a content problem long before it is a model problem. Duplicate versions, stale documents and wrong-entity content produce confidently conflicting answers — and that is a brand and compliance risk, not a search-tuning task.
- Classify every document by domain and legal entity on ingestion
- Extract with structure preserved — tables survive as tables, not flattened text
- Quarantine wrong-entity and low-confidence documents instead of publishing them
- Surface near-duplicates for a person to action — client content is never auto-deleted
- Propose FAQ and rule drafts from the document, for human approval
- Publish only after review — nothing reaches the corpus automatically
Change without a release
The administration platform
Prompt management
The assistant's instructions are versioned, two-person-approved and reversible — changed through the console, with no software release.
Rules management
Calculation parameters — LTV tiers, buffers, FOIR, credit bands — are administrator-editable, schedulable for a future date, versioned and two-person-approved.
Provisioning & budgets
Premium capabilities ship ready but off. One screen per capability: readiness → enable once provisioned → monthly budget ceiling. Cost is incurred only on purpose.
A configuration fault can never silently change a rule or interrupt the assistant — it falls back to a known-good default.
In practice
What a governed answer looks like
Illustrative lending scenarios — each answer computed by the rules engine, cited, and carrying the standard disclaimer.
| The user asks | What Ask Tara does |
|---|---|
| "For a ₹4,00,000 gold loan, what is the LTV?" | Returns the tier percentage from the administrator-maintained LTV table — computed, not recalled |
| "I have gold worth ₹3,00,000 — how much can I get?" | Returns the slab limit, and states how much more collateral would reach the next slab |
| "₹8L collateral, ₹4.2L outstanding, ₹5L scheme cap — top-up?" | Returns the lowest of the three limits, showing which constraint binds |
| The same question, asked in Hindi | The same exact figure, in Hindi — the computation is language-independent |
| "How do I reach Arun?" with a typo | Matches the same approved FAQ and returns it verbatim |
| A lending "top-up" question | Never returns HR or insurance content — domain separation is enforced at retrieval |
| Anything with no reliable source | Declines, asks for the missing detail, or routes to a person |
Security & compliance
BFSI-grade, built in
- Entra ID SSO — MSAL on the client, JWT bearer on the API; production-aligned identity from day one
- Server-side access control — retrieval is trimmed to what the user is entitled to see, enforced in the service, not the prompt
- Managed Identity and Key Vault — no static secrets anywhere in the deployment
- Data residency — Central-India deployment for DPDP and RBI alignment; the architecture is residency-gated by design
- Personal-data redaction — PAN and Aadhaar patterns stripped before anything is stored
- Runtime guardrails — a catalogue of behavioural controls on the live request path, each traced to a threat-model identifier
- Audit and rollback on every regulated change, with two-person approval
Architecture
What it is built on
- API
- ASP.NET Core (.NET 9) Web API with SSE streaming
- Front end
- React + Vite + TypeScript, with MSAL authentication
- Agent layer
- Microsoft Foundry via the Responses API — SME-tunable prompts, thinner code
- Retrieval
- Azure AI Search — hybrid BM25 plus vector, department-scoped indexes
- Embeddings
- text-embedding-3-large (3072 dimensions)
- Identity
- Entra ID SSO — MSAL.js and JWT bearer
- Persistence
- PostgreSQL (with pgvector for semantic cache)
- Ingestion
- Delta SharePoint indexer with retry and resilience
- Connectors
- REST and MCP protocol drivers; instances as admin configuration
- Hosting
- Azure Static Web Apps (SPA) + Container Apps (API)
- Channels
- Web, embeddable SDK for host applications, Teams tab, voice
Quality engineering
An assistant you can regression-test
"It seems better" is not an acceptance criterion. Ask Tara ships with an evaluation harness that treats the assistant as a black box and asserts on the things that actually matter.
Golden sets per domain
SME-editable question sets held in version control, one per department.
Citation assertions
Did the answer cite the document it should have cited?
Latency budgets
Every question carries a time budget the run must respect.
CI merge gate
The eval runs against a live API before a change is allowed to merge.
Feedback closes the loop. Thumbs up and down are captured on every answer, and the answer itself is stored with personal data redacted — so "what people asked" shows the question, the response and the outcome, and look-alike questions cluster into FAQ candidates.
Where it fits
Who deploys Ask Tara
NBFCs & lenders
Branch and contact-centre staff asking product, eligibility and process questions — where the answer is often a computed figure and always a compliance statement. Currently deployed inside a listed Indian NBFC across lending, HR, IT, operations and compliance knowledge.
Insurers & banks
Policy, claims and servicing knowledge with strict entity separation, audited change control and the ability to prove which version of a disclosure a user was shown.
Large regulated enterprises
Any organisation with tens of thousands of employees, a sprawling document estate, and an obligation to be right — HR policy, IT service desk, operations SOPs and compliance guidance.
Honest scope. Ask Tara is an assistant, not a decision system. It does not underwrite, approve or transact on its own authority. Write actions always pass through an explicit human confirm step, and the computation engine returns figures from administrator-approved parameters — the model presents them, it never derives them.
Point it at your hardest questions
Bring the questions your service desk answers badly and the policy documents nobody trusts. We will show you what a grounded, cited, computed answer looks like.