Skip to content
Get a live demo

Retrieval-augmented answer engine

The shortest distance between a question and a confident answer.

Zilzul drops into any product or help center and resolves 81% of user questions in under 1.4 seconds — so your team stops answering the same questions twice and your users stop getting lost in docs.

First-touch resolution
81%
Median answer latency
1.4s
Median contract to live
9days

30 minutes · with a solutions engineer · no sales script

Zilzul answering a product question with inline citations
  • BM25 + dense + reranker
  • p50 = 1.38s · p95 = 2.1s
  • Citation 1 · 2 · 3

Powering in-product assistants at 480+ teams · 1B+ queries served

  • Mercury
  • Ramp
  • Vanta
  • Linear
  • Notion

Ranked #1 in the G2 Spring 2025 Grid for AI Customer Support, ahead of Intercom Fin and Ada.

The mechanism behind the numbers

A hybrid retrieval stack tuned per tenant — not a one-size dense model.

The 81% resolution rate doesn't come from a clever prompt. It comes from a retrieval pipeline that's rebuilt for every customer's corpus: lexical and dense signals fused, then re-ranked by a model trained on enterprise Q&A. The result is an answer that cites the exact source span, not a generic "learn more" link.

See the benchmark methodology
  • 01

    Hybrid retrieval

    BM25 + dense embeddings run in parallel against the tenant's indexed docs, tickets, and product surfaces.

  • 02

    Per-tenant reranker

    A cross-encoder reranker is fine-tuned on the customer's own resolved tickets, not a frozen public model.

  • 03

    Grounded generation

    The model answers only from retrieved spans. If the evidence isn't there, Zilzul says so — and routes to a human.

  • 04

    Citation on every answer

    Every response links to the source span — paragraph and section — so reviewers and end users can verify in one click.

What you actually get in production

Four capabilities a serious buyer evaluates — each pinned to a number.

  1. 01

    Citations baked in by default

    Every response links to the source span — paragraph and section. No "learn more" hand-waves. Reviewers, auditors, and end users can verify every answer in one click.

    Median 3.1 cited sources per answer

  2. 02

    Per-tenant tuning, not a frozen model

    A reranker is fine-tuned on your own resolved tickets. Customers see a measurable lift versus the off-the-shelf baseline within the first two weeks of production traffic.

    +9.2 pts resolution over out-of-the-box dense retrieval

  3. 03

    9-day median from contract to live

    No fine-tuning, no model training, no six-week integration project. Connect a corpus, point Zilzul at your product surface, and ship the widget. Median across 2024 deployments.

    P90 = 21 days · P10 = 4 days

  4. 04

    SOC 2 Type II, ISO 27001, HIPAA-ready

    Compliance shipped in 2024. The only vendor in our tier with a public red-team report covering prompt injection, source poisoning, and cross-tenant leakage.

    Public uptime 99.98% · trailing 12 months

A 9-day deployment, told straight

What the median contract-to-live path actually looks like.

Most evaluators we talk to assume a retrieval deployment is a quarter-long integration. It's not — and the gap between expectation and reality is usually where deals die. Here's the day-by-day path our median customer walks, with the names of the artifacts they leave with.

Day 1 is kickoff. Your solutions engineer and a forward-deployed engineer join a shared channel. You send us the corpus — help center, Notion exports, Zendesk macros, in-product copy — and we start an initial index overnight. By Day 3 you'll have a private staging URL answering questions against your real content, with citations live. Days 4 through 6 are the tuning loop: your team flags wrong answers in a shared sheet, we adjust the reranker on the spots that matter, and we re-cut the eval set together. By Day 7 we're load-testing your traffic shape. Day 9 is the production cutover behind your feature flag, with a rollback switch and a runbook. If we miss the median, we tell you on Day 5 — not Day 30.

The reason this is fast is structural. Zilzul doesn't fine-tune a base model on your data; it indexes your data and runs a hybrid retrieval pass against it. There's no training job to wait for, no GPU cluster to provision, no model card to write. The 9-day number isn't a sales line — it's a measurement we publish quarterly in our customer letter, and it's the median across 184 production deployments in 2024.

If that sounds lower-risk than what you've been quoted, the demo is the place to pressure-test it.

One next step

Book 30 minutes with a solutions engineer.

We'll walk through your corpus, sketch a tuned configuration against your real questions, and give you a credible deployment date. No sales script, no follow-up sequence — just a working session.

First-touch resolution
81%
Median latency
1.4s
Contract to live
9days
Get a live demo

[email protected] · +1 (415) 555-0117 · San Francisco