AnomalyArmor
A data observability product, in private beta, that uses its monitoring signals to write and maintain a per-table markdown knowledge base for AI agents. Its blog argues this “living knowledge base” beats retrieval-augmented generation (RAG) for giving agents warehouse context.
The approach
AnomalyArmor builds one markdown knowledge base per monitored table, assembled from domain sections. The post says nine domain generators write them and names six:
- Schema drift — 90 days of column changes
- Freshness — update patterns, SLA violations and learned schedules
- Lineage — dbt-backed
- Data quality — metrics and validity checks
- Alerts — rules and history
- Context — tags, job history, git changes and an LLM-written summary
Nobody writes the files. Each domain carries a timestamp; when the underlying data changes, that domain is marked stale and regenerates the next time the knowledge base is read, while fresh domains are served as they are. The company says it has used the approach since late 2025, and describes it as a variation on Google’s Open Knowledge Format rather than a literal implementation of the draft v0.1 spec.
Why not RAG
The post’s reasons:
- Retrieval flattens structure — foreign keys and lineage edges become text fragments that lose their connections (“the chunks lie about structure”)
- A hidden failure point — a wrong answer could come from bad reasoning or bad retrieval, so there are two suspects instead of one
- The facts are already known — monitoring already detects when facts change, so embedding and searching on every query is wasted work
Its summary of the trade-off: “RAG’s hard problem is retrieval quality. A knowledge base’s hard problem is staleness.” RAG remains the better fit for large, changing, unstructured corpora such as support tickets, chat history and PDFs. A token-budgeting step still trims the knowledge base before model calls, because even a curated one can exceed a context window.
Limits
- The post names no author or publication date, and gives no benchmarks, accuracy, latency or cost figures; its comparison is qualitative
- It describes no human review step; the files are generated, not edited
- The approach depends on the metadata AnomalyArmor already collects through its warehouse connections
- AnomalyArmor is in private beta, and the post gives no pricing or availability beyond that
Related
- Open Knowledge Format and OKF vs Apache Ossie
- GraphRAG — a retrieval approach that keeps relationships
- Monte Carlo, Soda and Great Expectations — other data quality and observability tools
- Data governance for SMEs