Hi everyone,

I've been working on a knowledge retrieval system that can create distributable applications for several months and I'd love to share it with you!

The issue I've found with asking LLMs for advice is that they are trained on a large corpus of text but don't have much discernment. It's even worse when you ask a question like "What is Agile?" Agile is an emergent property of strong leadership and a compatible organisation design, not a methodology to shove down people's throats. However, an LLM is more likely to trend towards methodology-shoving answers. Asking for business advice is likely to result in what a March 2026 Harvard Business Review study called "trendslop." (https://hbr.org/2026/03/researchers-asked-llms-for-strategic...) LLMs are good with language, not discernment, and more language does not equate to more truth.

I also needed a way to mentor globally distributed teams, especially when time zones made real-time support hard (I live in New Zealand and had clients in Europe, for instance). I wrote all sorts of guidance--like how to write a strong After-Action Review--and found that it was still challenging for folks to succeed with it.

Grounded Forge is a set of MD files and small helper scripts, run within either Claude Code (extensively tested) or Codex (just ported, smoke-tested). Give it a meaty source (the demo corpus is 27 open texts, 12 of which are full textbooks) and it creates multiple source-verified summaries in a 9-pass ingestion process while maintaining the original in an MD file. It also creates an extensive concept index in JSON using an LLM, with ChromaDB running as a backup. My daily use has shown this approach to be superior to letting the LLM simply grep the sources.

My own corpus has 282 sources; 178 of them have short summaries about how they inform AAR practice, 184 on retrospectives (the rest were judged not relevant to those tasks at ingestion). This means that when I ask an AAR question, it's able to see what 178 sources say only on the topic of AARs without having to sort that out at run-time.

The net effect of this is a RAG-like system, with virtually no vendor lock-in (the "code" is open-source and easily ported to other models/harnesses), that is simultaneously more efficient and thorough at runtime, with the tradeoff of increased costs at ingestion time.

The best way to see what it does is by example. Here's an annotated run of the prompt "/answer-from-corpus Please draw some interesting parallels between US Marines war-fighting doctrine and modern "textbook" business administration." https://github.com/chrisgagne/grounded-forge/blob/main/docs/.... This shows off the tooling's ability to draw distinctions and parallels between seemingly unrelated texts.

There are some limits: it costs about $1-5 to ingest a book-length source on API tokens with a frontier model. Citation behaviour is prompt-inherited, not build-enforced; no head-to-head benchmark against conventional RAG yet. If the sources are texts already in the training data, I've found that the models can often have a good sense of their conceptual contents just by looking at a directory's worth of file names.

The end result of this is:

* Dozens of development teams I've worked with now have an AAR assistant that is reproducibly grounded in the best evidence I could find for how to run solid AARs and it's improved their investigations

* I'm finding that I am able to research and write well-grounded articles, with each claim directly verifiable to the original source

* I'm able to mine seemingly disparate topics like Tibetan Buddhism and adaptive software development and draw interesting parallels that I could not before

It is my hope that if enough people find this useful, we can expand the open corpus so that the default tool out of the box can provide solid, grounded advice for more questions.

Show HN: Grounded-forge – RAG with summaries and task views precomputed ingest