Product · WeaveAI Anchor
Answers grounded in your own documents.
Anchor is retrieval-augmented generation over your company's documents. It answers from the sources you approved, cites the passage behind every answer, and refuses rather than hallucinating when the documents don't cover the question.
483
automated tests across the retrieval pipeline and the Anchor product — run on every change
100%
of answers either cite a retrieved passage or are refused; the pipeline strips citations that point at nothing
4
document sources you can connect today: Confluence, Google Docs, PDF, pasted text
Grounded by construction, not by prompting.
Three steps — connect your sources, retrieve instead of invent, cite or refuse.
Point it at your documents
Connect Confluence, Google Docs, PDFs, or paste text directly. Documents are chunked and indexed once — nothing needs reformatting, and your existing internal knowledge base stays where it is.
Every answer is retrieved, not invented
This is retrieval-augmented generation: the question goes to your indexed documents first, and only the passages that actually match are put in front of the model. Nothing is answered from training data.
Cited or refused — never guessed
Every answer carries a reference to the passage that supports it. If the documents don't cover the question, the system declines and lists what it checked, rather than producing a hallucination that reads like a fact.
Two examples, same boundary.
Different question shapes, same rule underneath: answer from the document, or say so.
Try the boundary
Pick a question. Two of them are answered by a robotics manufacturer's product and policy documents; two aren't. Scripted examples using Anchor's real response shapes — not a live query against a customer's data.
How long is the warranty on the R-40 arm?
The R-40 industrial arm carries a 24-month warranty from the date of commissioning, covering parts and labour for defects in materials and workmanship. [1]
[1] R-40 Warranty Terms · §2.1
“All R-40 series arms are warranted for twenty-four (24) months from the date of commissioning against defects in materials and workmanship, inclusive of parts and labour.”
Boundary: If the answer isn't in your approved documents, the system refuses instead of guessing. This is what stops a hallucination from reaching the person who asked, and it is the reason the answers can be trusted without checking every one.
Built for more than one customer at a time.
The parts that matter once this is running for real accounts, not a demo.
Fails safe by design
The refusal is enforced in the retrieval pipeline, not requested in a prompt. A citation pointing at a passage that was never retrieved is stripped from the answer before it reaches you, and an answer left with no citations at all is flagged as ungrounded.
Isolated per customer
Each tenant's documents live in their own data partition — never a shared index with metadata filtering to remember. One customer's data is structurally invisible to another's queries.
Hybrid retrieval, cost-routed
Keyword (BM25) and vector search run together, get fused with reciprocal rank fusion, and are reranked before generation. Simple questions route to a cheaper model automatically, so accuracy costs what it needs to and nothing more.
Works with what you already have
Product manuals, technical specs, policies, pricing sheets, compliance documents — whatever factual reference material your team already maintains, in the formats you already keep it in.
Priced per query, not per seat.
Start free, no card required — every plan, including Free, runs the real pipeline: hybrid retrieval, citation enforcement, and per-tenant isolation.
Free
$0 / mo
50 queries / mo
- 1 connected data source
- Full citation + refusal behavior — not a crippled demo
- No card required
Starter
$99 / mo
1,000 queries / mo
- Up to 5 connected data sources
- Hybrid retrieval — dense + keyword, fused by rank
- Cost-routed generation (cheap questions, cheap model)
- Per-tenant data isolation
Enterprise
Let's talk
custom query volume
- Unlimited connected data sources
- Dedicated namespace & deployment
- Custom source connectors
- Priority support
Prices in USD, excl. taxes — local taxes and currency are applied at checkout.
Questions teams ask before rolling this out.
Direct answers, no hedging.
What is retrieval-augmented generation (RAG)?
RAG is a technique where a language model answers from documents you supply rather than from what it memorised during training. The question is first used to retrieve relevant passages from an index of your own content, and only those passages are given to the model to answer from. That is what makes the answer checkable: there is a specific source behind it. Anchor is a RAG system with the citation and refusal rules enforced in the pipeline rather than requested in a prompt.
How does this prevent wrong answers?
Three things, in order. Retrieval limits the model to passages from your approved documents. Citation integrity checks every reference in the generated answer against the passages that were actually retrieved, and strips any that don't match. And if nothing relevant was retrieved, or the answer ends up with no valid citation, the system refuses rather than returning it. A hallucinated citation cannot survive that chain, because the check is code and not a prompt instruction.
What kind of documents can I use?
Product manuals, technical specifications, policies, pricing sheets, compliance documents — any factual reference material your team uses to answer questions. You can connect Confluence spaces, Google Docs, upload PDFs, or paste text directly. We help you identify which documents to include during setup.
Can I use this as an internal knowledge base assistant?
That is the most common deployment. Point Anchor at the Confluence space or document set your team already maintains and it becomes a document Q&A layer over it — the same content, reachable by question instead of by search and scroll. Because it cites the passage behind each answer, someone can check the source without leaving the answer, and because it refuses when coverage is missing, gaps in the knowledge base surface as refusals instead of being papered over with plausible-sounding text.
Can each of my business units or clients keep their own data separate?
Yes — each tenant gets its own isolated data partition, not a shared index with a filter that has to be remembered on every query. That isolation is structural, so a bug in a query can't leak one tenant's documents into another's answers.
Does this replace the assistant I already have, or work alongside it?
It's built to slot into an assistant or workflow you already run — the grounded, cited answers support the existing conversation flow rather than requiring a rebuild.
Start Now
Give your assistant a boundary it can't cross.
Connect a document, ask it something real, and see the citation attached to the answer.
Written and maintained by the WeaveAI engineering team · last updated 2026-08-15