HandbookBrain
A graph-database-based knowledge graph system for company rules that models documents, meetings, decisions, and implementation changes as a traceable, versionable graph, enabling agents to query the rationale behind rules and to perform impact analysis.
Timeline (last 90 days)
HandbookBrain maintains a versioned knowledge graph on FalkorDB that holds company documents, meetings and chats, merge requests for implementation decisions, and the systems, CRM fields, metrics, and dashboards the documents touch.
HandbookBrain uses sections rather than documents as the unit of versioning, creating a new Section node for each change, pointing to the old version via SUPERSEDES, and carrying valid_from/valid_to timestamps.
In HandbookBrain, an Answer node references via CITES edges the exact section versions it used, so that later changes can find the answers they break.
HandbookBrain extracts decisions from meetings, chats, or tickets, with speakers and verbatim quotes, links them to the relevant sections, and reconciles them against implementation commits by reading diffs.
HandbookBrain implements hybrid retrieval entirely within FalkorDB: a vector index of section embeddings plus a full-text index of titles and bodies, fused with reciprocal rank fusion (RRF) and filtered to the current version.
HandbookBrain's agent is a tool-calling loop wrapping 10 parameterized read-only Cypher tools; every assertion must carry references actually returned by the tools, which is checked by code.
HandbookBrain's reviewer runs on a different model, sees only the question, evidence, and draft, and can send it back for rewriting.
In the demo, the real MR !19600 on May 21 added "Trial Threshold Scoring" to a Marketo page, and HandbookBrain linked it to a synthetic marketing operations meeting from 13 days earlier.
On July 23, the real MR !20534 changed the trial threshold from 25 to 31, causing the June 1 answer to be judged outdated by the diff; the answer was refreshed to 31, with the meeting, decision-maker, and MR attached.
HandbookBrain's audit compared change dates, MR numbers, and version chains against git log, passing 2562/2562.
6 more timeline events
All Facts (16)
HandbookBrain maintains a versioned knowledge graph on FalkorDB that holds company documents, meetings and chats, merge requests for implementation decisions, and the systems, CRM fields, metrics, and dashboards the documents touch.
50%UnverifiedHandbookBrain uses sections rather than documents as the unit of versioning, creating a new Section node for each change, pointing to the old version via SUPERSEDES, and carrying valid_from/valid_to timestamps.
50%UnverifiedIn HandbookBrain, an Answer node references via CITES edges the exact section versions it used, so that later changes can find the answers they break.
50%UnverifiedHandbookBrain extracts decisions from meetings, chats, or tickets, with speakers and verbatim quotes, links them to the relevant sections, and reconciles them against implementation commits by reading diffs.
50%UnverifiedHandbookBrain implements hybrid retrieval entirely within FalkorDB: a vector index of section embeddings plus a full-text index of titles and bodies, fused with reciprocal rank fusion (RRF) and filtered to the current version.
50%UnverifiedHandbookBrain's agent is a tool-calling loop wrapping 10 parameterized read-only Cypher tools; every assertion must carry references actually returned by the tools, which is checked by code.
50%UnverifiedHandbookBrain's reviewer runs on a different model, sees only the question, evidence, and draft, and can send it back for rewriting.
50%UnverifiedThe author acknowledges evaluation limitations: the questions came from a single dataset, and the agent was modified after seeing it fail on these questions, so numbers would be lower on new questions.
50%UnverifiedThe evaluation found a real bug: a section named "Trial Threshold Scoring" was buried under the generic "MQL and Lead Scoring," causing the agent to answer 100 instead of 31, and the fix was to add title-matching candidates to the search.
50%UnverifiedHandbookBrain's audit compared change dates, MR numbers, and version chains against git log, passing 2562/2562.
50%UnverifiedHandbookBrain's audit compared decision-makers, locations, and quotes against meeting records, passing 21/21; compared document owners against CODEOWNERS, passing 122/122; and compared decision-to-implementation MR links against hidden tags, with 0 broken links.
50%UnverifiedIn an evaluation of 29 questions × 3 trials, HandbookBrain scored 87% overall, RAG + git log scored 66%, and plain RAG scored 59%.
50%UnverifiedOn "name the MR that implemented the decision," HandbookBrain was correct 18/18, RAG + git log only 4/18, and plain RAG 0/18.
50%UnverifiedOn "old/new values of changed rules," HandbookBrain was correct 15/15, RAG + git log 9/15, and plain RAG 6/15; on meeting dates all three were 18/18; on plain lookup all three were 70%.
50%UnverifiedIn the demo, the real MR !19600 on May 21 added "Trial Threshold Scoring" to a Marketo page, and HandbookBrain linked it to a synthetic marketing operations meeting from 13 days earlier.
50%UnverifiedOn July 23, the real MR !20534 changed the trial threshold from 25 to 31, causing the June 1 answer to be judged outdated by the diff; the answer was refreshed to 31, with the meeting, decision-maker, and MR attached.
50%