Knowledge base

Upload a PDF.
Your AI answers from it minutes later.

Drop in product docs, help articles, warranty policies, anything. We chunk it, embed it in Qdrant, and your AI starts answering with citations back to the exact page. No prompt engineering, no retraining.

12+ file formats supported
200MB max file size
< 400ms p95 query latency
product-handbook-v3.pdf 3.2 MB Indexed
shipping-policy.md 42 KB Indexed
pricing-2026-q2.xlsx → pdf 1.8 MB Indexed
help.example.com/faq web Indexed

PDFs, websites, Notion pages — you name it

Drop in what you already have — no re-formatting, no exports to some proprietary shape.

PDF
.pdf · .pdf/a
Per file200 MB · OCR for scanned
Product manuals, handbooks, warranty docs, contracts — pretty much anything you email clients.
Markdown
.md · .mdx
Per file 10 MB
Perfect for engineering wikis, changelogs, and anything you already write in Markdown.
Notion
notion.so/...
Workspace export unlimited pages
Connect a workspace and sync selected pages — updates land in your KB automatically.
Web pages
https://...
Site crawl 500 pages · refresh daily
Your help center, docs site, or blog — we crawl and re-crawl on a schedule you set.
Word / Excel
.docx · .xlsx
Per file 50 MB
Your sales playbooks, price sheets, handbooks — upload them as-is, tables included.
Plain text
.txt · .csv · .log
Per file 20 MB
Exports from other tools, FAQ dumps, chat transcripts — the lowest-friction format.
API / JSON
POST /v1/kb/ingest
Per request 2 MB · auto-batched
Sync from your CMS, wiki, or internal tools — one endpoint, idempotent upserts.

Upload to answering in four steps

From dragging in a PDF to a live AI answering questions from it — usually under five minutes.

01
Drop in your sources DROP

Drag a folder of PDFs, paste a help-center URL, or connect your Notion workspace. Mix and match.

Drop files here PDF, Markdown, Word, Excel, Notion, URL
02
We chunk and embed AUTO

We extract text, OCR images, split into semantic chunks, and embed everything in Qdrant. You don't touch any of this.

Extracted 184 pages from product-handbook-v3.pdf
OCR'd 12 tables and 6 diagrams
Embedding chunk 312 of 506...
Indexing to Qdrant · est. 40s remaining
03
Test it right there TEST

Ask questions in the dashboard and see exactly which chunk the AI pulled. Good answer? Ship it. Weird answer? Keep reading.

What's the warranty period on the Velvet 2?
Standard warranty is 24 months, covering defects in materials and workmanship. Wear from normal use isn't covered.
product-handbook-v3.pdf · p.47
04
Correct what's wrong FIX

Didn't love an answer? Thumbs-down, leave a note, upload a better source. The AI gets smarter with every correction.

Shipping to Canada usually takes 7–10 business days.
Not quite — numbers are from 2024
Fixed: added "canada-shipping-2026.md" to knowledge base

Sources stay visible

Every answer shows which file it came from. If nothing matches, the AI says so instead of making things up.

What's actually in the Velvet Latte?
AIMei · barista
The Velvet Latte uses oat milk as the base, with a pinch of Himalayan salt and a finish of Tahitian vanilla. No artificial flavors. velvet-ingredients.pdf · p.2
Can I get it with oat milk?
AIMei · barista
Yes. Any drink on the menu can be swapped to oat, almond, or soy milk at no extra charge. Just say the word when you order. barista-handbook-v7.pdf · p.14
Do you sell the beans by the kilo?
AIMei · barista
I don't have info on that one yet. Want me to loop in a human, or you can ask our team directly. No source found — handoff offered

Retrieval-augmented, not memorized

When someone asks a question, we search your docs for the most relevant chunks first. Those chunks go to the AI along with the question — so the answer comes from your content, not from training data.

If nothing matches above the confidence threshold, the AI says it doesn't know. No hallucinated specs, no made-up prices, no "I think maybe..."

Every response shows which chunk it used. Click the citation and you jump straight to the source — page, section, line.

Under the hood: Qdrant vector search, hybrid semantic + keyword, p95 under 400ms across the board.

What the AI said
The Velvet Latte is built on oat milk, finished with a pinch of Himalayan salt and real Tahitian vanilla. Everything on the menu can be swapped to almond or soy if oat isn't your thing.
Generated in 1.4s · 2 chunks retrieved · confidence 0.87
What the doc actually says
Velvet Latte (signature) Base: oat milk (Oatly Barista) Accent: Himalayan pink salt, Tahitian vanilla extract Milk swaps: almond, soy — no surcharge Allergens: contains oats; vanilla is alcohol-based
velvet-ingredients.pdf · page 2, chunk 14

Manage it like real content

Docs go stale. Your knowledge base shouldn't. Tools to version, update, and audit — built in.

Version history

Every upload is versioned. Roll back in one click if a new doc broke something, or keep both versions live for A/B testing.

v3.2 current 6 min ago
Added Q2 2026 pricing, updated warranty terms
v3.1 Added shipping policies for EU and UK 3 days ago
v3.0 Initial handbook upload 2 weeks ago

Incremental uploads

Add new docs any time without rebuilding the whole thing. Only the changed chunks get re-embedded, so costs stay low and nothing goes offline.

  • Upload a single doc or a whole folder — incremental, not full rebuild
  • Diff-aware: if only page 3 changed, only page 3 gets re-embedded
  • Scheduled re-crawls for URL sources — daily, weekly, or on demand
  • Webhook on ingest complete, so you can wire it into your CI

Delete & review

Mis-uploaded the wrong file, or a doc is flat-out wrong? Pull it out, and every cached answer referencing it stops serving.

  • One-click delete pulls a doc and its chunks from search in seconds
  • Old answers citing that doc are flagged so you can review them
  • Soft delete keeps it around for 30 days in case you change your mind
  • Audit log shows who deleted what, and when

Improvement loop

Thumbs-down an answer, leave a note, and your team gets a ticket to fix the doc or add a missing chunk. The next customer gets a better answer.

Customer asksquestion in chat
AI answerswith citation
Team improvesupdate the doc
Feedbackthumbs up/down

Things people ask before uploading

How long does indexing take?
Anywhere from 10 seconds for a short Markdown file to a couple minutes for a 200-page PDF. You'll see live progress in the dashboard, and we email you when it's done if you step away.
Does it handle images and scanned PDFs?
Text does. Images and tables get OCR'd automatically — scanned PDFs, screenshots, the works. Handwriting is hit-or-miss. If OCR confidence is low on a chunk, we flag it so you can check.
Does it pull updates when my source doc changes?
Yes. URL sources can be set to re-crawl daily, weekly, or on demand. For file uploads, you upload the new version and the old one rolls to an archived state — answers automatically switch to the new version.
What if the AI can't find the answer?
If the answer isn't in any doc, the AI says so and offers a human handoff rather than making something up. That's the whole point of citations — they show work. You can also set a confidence threshold and have the AI defer to a human below it.
Who sees my uploaded documents?
Your docs are encrypted at rest and in transit. Embeddings run in our Qdrant cluster — we don't use your content to train anything. Details in our privacy policy.
What languages does it support?
English, Chinese (simplified + traditional), Japanese, Spanish, French, German, Portuguese, and a dozen others. You can ask in one language and the AI will pull from sources in another — citations still point to the original.
How is this different from Roleplay AI?
Think of it this way: Knowledge is the what — facts, specs, policies. Roleplay is the how — tone, personality, what the AI sounds like. You need both: the persona without knowledge sounds confident but wrong; knowledge without a persona sounds like a search engine.
Can I have different docs for different personas?
Absolutely. Different personas can read from different knowledge bases — your sales bot might see the pricing doc your support bot doesn't. Set it up in persona settings .

Pairs well with

Knowledge base is one piece. Here's how it wires into the rest of Lane.Chat.

Turn your docs into a coworker that actually answers.

Drag in a PDF, paste a URL, watch your AI start answering with real answers.