Skip to content

Start typing to search products, founders and collections.

↑↓ navigate↵ openesc closeFull search →
Anchor logo

Anchor

Verified

An AI research assistant that shows its sources or says it does not know

Launched Week 33, 202624k viewsFreemium · from $24/mo
641Visit site
Anchor interface — main workspace
  • Anchor interface — detail view
  • Anchor interface — settings

About Anchor

Anchor answers research questions against your own document corpus and cites the exact passage behind every claim. When the corpus does not contain an answer it says so instead of generating a plausible one — which is the only behaviour that makes an AI research tool usable for work that matters.

Features

  • Passage-level citations

    Every claim links to the exact paragraph behind it, not just the document.

  • Honest misses

    Returns "I don't have that" rather than falling back to model knowledge.

  • Self-hosted embeddings

    A corpus under NDA never leaves your network.

  • Confidence floor

    A tunable threshold below which Anchor declines to answer at all.

  • No persona

    A research interface, not a chatbot. Query in, cited answer out.

  • Corpus diffing

    Shows which answers changed when you add or remove documents.

The story

1 min read

The problem is not fluency

Language models are already fluent enough. The failure mode that matters for research is that a confident wrong answer costs more than no answer at all, and you cannot tell them apart by reading.

Anchor is built around making that distinction visible.

How it works

Every claim in an Anchor answer is anchored to a span in a source document — not a document-level citation, a passage-level one. Click any sentence and you land on the exact paragraph it came from.

When retrieval returns nothing above the confidence floor, Anchor returns "I don't have that" and shows you what it did find. It does not fall back to the model's parametric knowledge, because a tool that silently switches between "from your documents" and "from the model's training data" is worse than useless for compliance, legal or technical research.

What it is not

Anchor is not a chatbot. There is no persona, no follow-up small talk, and no attempt to be conversational. It is a research interface: query in, cited answer or an honest miss out.

It is also not a general-purpose assistant. It only knows what you have given it.

Deployment

Your documents are embedded and stored in your own infrastructure. Anchor supports self-hosted embedding models, which means a corpus under NDA never leaves your network — the usual blocker for teams who would otherwise have adopted something like this two years ago.

Pricing

Free up to 500 documents. Self-hosted embeddings on all plans.

Free

Free

  • 500 documents
  • Passage-level citations
  • Self-hosted embeddings
  • One user
Popular

Team

$24/month

  • Unlimited documents
  • Shared corpora
  • Confidence tuning
  • Corpus diffing
  • SSO

Enterprise

Custom

  • Air-gapped deployment
  • Custom retrieval pipeline
  • Audit log
  • SLA

Changelog

  1. v1.2

    Corpus diffing

    Adding or removing documents now shows which previously-answered questions changed, so you can see when a policy update invalidated an earlier answer.

Frequently asked questions

Does Anchor hallucinate?

Every claim is anchored to a retrieved passage, and when retrieval returns nothing above the confidence floor Anchor declines rather than generating. That eliminates the specific failure of inventing facts from parametric knowledge. It cannot eliminate misreading a passage it did retrieve, which is why the citation is always one click away.

Where is my data processed?

Embeddings can run entirely on your own infrastructure using a self-hosted model, on every plan including the free tier. In that configuration your documents never leave your network.

What file types are supported?

PDF, DOCX, Markdown, HTML and plain text, plus Notion and Confluence via read-only connectors. Scanned PDFs are OCR'd on ingest.

How large a corpus can it handle?

Comfortably into the low millions of passages. The free tier caps at 500 documents; paid plans are limited by your own storage rather than by us.

Can I use it as a chatbot?

You can ask it follow-up questions, but it will not maintain a persona or answer anything outside your corpus. That is deliberate — the moment a tool blurs "from your documents" and "from training data", its citations stop meaning anything.

Anchor alternatives

Similar products worth comparing before you commit.

  • Quill

    A writing environment where the outline is a real object, not a heading hierarchy

    If you are writing the document rather than searching it.

    ProductivityOne-time
    503

  • Driftwood

    A local-first sync engine you can drop into an app you already shipped

    Shares the commitment to keeping data on your own infrastructure.

    Developer ToolsOpen source
    1.5k

Works with

Discussion93

Markdown supported. Be useful, not promotional.

Related products

Ranked by shared categories, tags and explicit alternative relationships.

  • Quill

    A writing environment where the outline is a real object, not a heading hierarchy

    ProductivityOne-time
    503

  • Halden

    A self-hosted CRM for consultancies that will not put client data in the cloud

    SaaSOpen source
    588

  • Northlight

    Featured

    Design tokens that stay honest between Figma and production code

    DesignFreemium
    981