Better inputs. Better agents.

Convert any file to Markdown before it reaches the model.
Read what comes back, annotate, and send back.

Or drop any file anywhere on this page

Retrieval pipeline

The indexer runs on every commit and writes chunk embeddings to the vector store. Chunk size is fixed at 512 tokens, which the retrieval layer assumes downstream.

MayaStill 512? Ingest says 256.

Queries are rewritten before search. The rewriter shares a cache with the answer model, so a cold start costs two round trips instead of one.

Rerank top 50 candidates with the cross-encoder.
Drop anything below 0.42 similarity.
Return citations with byte offsets.

Offsets are stable only while the source file is unchanged.

Evaluation runs nightly against the golden set. Regressions block the deploy unless a maintainer overrides.

MayaStill 512? Ingest says 256.

Note for the agentCache sharing is out of scope for this PR — split it out.

A reading surface for documents written by agents. Markdown is where specs, plans, and agent output already live. Annotate makes it reviewable by people and returnable to the tools that wrote it.

48-page PDF96,400

→

.md50,100

CONVERT

Convert to Markdown

Any supported file becomes Markdown before it reaches the model. A 48-page PDF that costs 96,400 tokens as page images costs 50,100 as Markdown.

L14This line is wrong.

ANNOTATE

Comment on the exact line

Read agent output as a document instead of scrolling a raw blob. Highlight the line that's wrong and leave the note on it.

architecture.notes.mdL14 "512 tokens"Still 512? Ingest says 256.

RETURN

Send the notes back

The agent gets a Markdown file with your highlights and notes anchored in place, not the document pasted again. It knows which lines to change, so it takes one revision instead of three.

Reduce cost on every upload

Save thousands of tokens by converting heavy files into Markdown without leaving Claude.

Install the plugin and every attachment is cost-optimized before it reaches the model.

Share document

getannotate.dev/d/architecture-md

Any file as Markdown

Convert PDFs, Word docs, images, and HTML to Markdown before you send them to an agent.

Same content, fewer tokens.

Conversation

Take the three highlighted lines and rewrite the chunking section to match ingest.

Updated the section to 256 tokens and noted the downstream assumption. Two of your notes are resolved.

architecture.md · revision 4

Connect over MCP

Your agent reads the notes itself over MCP — no copying comments into a prompt.

It gets the highlight and the lines it points at.

Repository

docsarchitecture.md4 notesretrieval-eval.mdrollout.md

annotate/review-notes → main

Speed up your feedback

Drop any file to review

PDFs, Word docs, and HTML convert to Markdown. Renders instantly.