Backdraft

Backdraft creates shareable, self-contained documents with clickable citations, inspired by my work on financial AI at HighRoad. It’s shipped as a skill and a CLI, creating a general toolkit for use with any agent. Point it at your files, and your agent is verifiably factually-grounded.

HighRoad’s screening product resolved every extracted figure and data point in a generated investment memo back to the page or the spreadsheet cell it came from. Untraceable or unsourced generations are effectively useless in finance where small mistakes add up in a 100M loan.

The shared token database

One database sits between the source documents and the writing, and every part of the system speaks its identifiers.

1. Every span the writer sees becomes a minted token, so the database holds a record of what was read.
   • Source documents reach the writer through the read and search commands.
   • If an agent did not actually see a citation or invents one, the system calls it out.
2. Anchors take their identity from the document’s content, ensuring citations survive reprocessing.
   • The token is the file’s bytes plus a location inside it rather than a row created by a run, so re-ingesting the same document returns the same tokens.
3. Each anchor stores the cited evidence, so the document is portable without the database behind it.
   • The verbatim snippet, the content hash, the location, and the page image the model read all travel with the claim.
4. Every citation that fails to resolve appears in the report.
   • Agents have the chance to correct on their own or with user direction.

Backdraft is a SQLite file in a hidden directory and four commands, ingest, read, bind, and render. It produces one self-contained HTML file that opens offline on any machine. One click claim attribution is now local and portable.