Why we built Anserra: RAG chatbots that don’t hallucinate

Anserra started because too many chatbots were built to sound right instead of be right.

The problem we couldn’t stop thinking about

We work with businesses across support, e-commerce and internal tooling. Every single one of them had tried a chatbot at some point. Almost all of them had turned it off.

The failure mode was consistent: the chatbot answered confidently, was frequently wrong, and there was no way to prove which document produced any given answer. The support team stopped trusting it. Customers noticed. It went back in the drawer.

Meanwhile the underlying technology — retrieval-augmented generation — was proven. The problem was product decisions, not model capability. Every mainstream chatbot vendor was optimising for time-to-first-answer instead of trustworthiness of answers.

We wanted the opposite. A chatbot you can point at your content and trust to (a) answer from it, or (b) refuse. Nothing in between.

The three product bets

Bet 1: Every answer must show its source, by default. Not “citations available.” Every answer. If a claim isn’t backed by a passage from your knowledge base, the chatbot didn’t retrieve it — the model made it up. Anserra doesn’t ship a mode where citations are hidden.

Bet 2: Hybrid retrieval is the baseline, not an upgrade. Semantic search is astonishingly good at paraphrase and blind to exact strings. Real customer questions include SKUs, error codes, order references, model numbers — strings whose meaning is the exact characters. We built hybrid retrieval into the default query path from day one.

Bet 3: Nothing runs on a black box. Postgres for storage. pgvector for embeddings. Open standards for the API. If a customer wants to leave, they can export everything. If they want to inspect what we do with their data, they can read the source of the ingestion pipeline — most of it is in our public documentation.

Every product decision points back to one of these three.

What the first 100 customers taught us

Six months in, here’s what we learned that we didn’t expect:

People love the “I don’t know” reply. We assumed customers would push back on a chatbot that refuses. Instead, the first month of support tickets was overwhelmingly positive about this specific behaviour. A chatbot that refuses is a chatbot they can trust with real customers.

Content gaps are the second product. The Analytics tab surfaces questions the chatbot couldn’t answer well. That list turns out to be the single most valuable output of the tool. Teams close it weekly and their help centres get measurably better. The chatbot is doing content-strategy consulting as a side effect.

Agencies want multi-tenant workspaces urgently. We didn’t build for agencies initially. Within two months, agencies were our fastest-growing segment. They resell white-label chatbots to their clients from a single workspace. We rebuilt the workspace model around this.

Embedding is where friction lives. The chatbot works. Getting it onto a customer’s site is where most trials died. We wrote per-platform snippets (Shopify, WordPress, Webflow, Next.js, plain React) and time-to-first-live dropped from ~2 hours to ~5 minutes.

Honesty is a moat. Vendors that promise “no hallucinations” and then hallucinate are the norm. Being explicit about what our product doesn’t do turned out to be a stronger sales motion than promises. Customers appreciate it, and it’s much easier to sustain than the alternative.

What’s next

Voice. The same retrieval-first architecture works for phone calls. It’s harder — real-time streaming, longer context management — but the same product bets apply.

Better integrations. We’ve focused on the widget and the API. Slack, WhatsApp, Telegram, Intercom hand-off — all queued.

Model diversity. OpenRouter unlocked this. We can offer per-chatbot model choice — a customer can pick Claude for a nuanced support bot and Gemini for a factual lookup bot — without maintaining N provider integrations.

Better developer tooling. A CLI, JSON export/import, staging environments. Our REST API is complete; the tooling around it is under-built.

The company

Anserra is a product of Onzira Labs, an AI automation and software company based in New Delhi. We build software for a living — enterprise SaaS, industry-specific tools, custom AI agents. Anserra is the chatbot product we always wished existed for our own client work.

If any of this resonates, try it free. No credit card. Live in about five minutes. If it doesn’t fit, tell us why — we listen, and we still iterate faster than most companies our size.

Thanks for reading.