Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.stratta.ch/llms.txt

Use this file to discover all available pages before exploring further.

TreeRAG is the reasoning model behind Stratta. It explains why answers come with citations you can open and verify, rather than approximate references.

RAG, briefly

“RAG” (retrieval-augmented generation) means giving an AI the right source text before it answers, so it reasons from documents instead of memory. The question is how you retrieve the right text. Stratta’s answer is TreeRAG.

The usual way: vector embeddings

The common approach splits a document into many small chunks, converts each into a numeric vector, and at query time retrieves the chunks whose vectors are closest to the question’s vector. It’s fast and general, but for engineering norms it has two costs:
  • Citations get fuzzy. A chunk boundary rarely lines up with a clause or page, so the reference back to the norm is approximate.
  • Structure is lost. Norms are hierarchical and cross-referenced. Flat chunks discard the chapter → section → clause tree that tells you where a rule lives.

Stratta’s way: navigate the tree

Stratta stores each norm as its actual table of contents — a tree of chapters, sections, sub-sections, and annexes, inspired by the PageIndex approach. There are no embeddings. Claude navigates that tree like an engineer with the printed norm:
StepWhat Claude doesTool
1See which norms existlist_norms
2Read the chapter structureget_toc
3Drill into a chapterget_subtree
4Read a section’s full contentget_section
5Search by keyword when the path is unknownsearch_in_norm
6Follow links to other normsget_cross_refs
7Open a diagramget_figure
Each section retains its exact path and page range. So when Claude cites [SIA 261, 8.2, p. 44], that location is real and openable — what we call a deterministic citation.

Why cross-references matter

SIA standards deliberately distribute a single rule across several norms: the action in SIA 261, the safety factors and combinations in SIA 260, the material rules in the domain norm (concrete in 262, steel in 263…). Stratta records these links explicitly so Claude can follow them. Missing cross-references is the most common cause of incomplete answers — TreeRAG is built to avoid it.

Trade-offs

TreeRAG is excellent for structured, citation-critical documents like norms. It does require a high-quality table-of-contents tree, which is why ingestion is careful and agent-driven (see Ingest a norm). For unstructured prose, plain embeddings can be simpler — but that’s not the problem Stratta solves.

Architecture

Where the tree is stored and served.

MCP read tools

The tools that walk the tree.