What retrieval missed: a company brain on FalkorDB, tested on 1,203 held‑out questions
Описание и применение
Модель поиска пропустила контекстные связи в корпоративных данных, тогда как граф‑тул на FalkorDB вернул все релевантные данные, улучшив полноту ответов.
Задача
Retrieval models often miss contextual relationships in enterprise data, leading to incomplete or incorrect answers.
Результат
The graph‑tool agent retrieved all ten relevant employee IDs for a spec question, whereas retrieval returned only one. In a held‑out question about bug fixes, the agent correctly identified Julia Miller as the top contributor, matching the gold answer, while retrieval named four other engineers.
Рабочий процесс
- 1. Build a FalkorDB graph per product from structured records.
- 2. Define nine read‑only Cypher tools for querying the graph.
- 3. Deploy an agent that calls these tools to answer questions.
- 4. Run the agent on 1,203 held‑out HERB questions.
- 5. Compare results with retrieval‑based answers and gold answers.
Доказательства
I built inherit‑brain for Graph Hacks, Track 3: Company Brain.
The agent made one tool call (doc_people), which ran three read‑only queries in 21.9 ms. It returned ten people.
Julia Miller had two fixing PRs across the four matched bugs, which gave HERB's gold answer.
- Дата записи: 2026-10-18T06:41:16.527Z