tenfingers/powershell-docs-rlat

powershell-docs — rlat knowledge model (v2.0)

A Resonance Lattice knowledge model of MicrosoftDocs/PowerShell-Docs at commit 58137d73, scope reference.

Quick start

pip install rlat
huggingface-cli download tenfingers/powershell-docs-rlat powershell-docs.rlat --local-dir .
rlat search powershell-docs.rlat "your question" --top-k 5

The model uses remote storage mode — passages reference source files at raw.githubusercontent.com pinned to the commit SHA above. The first query fetches each cited source once and caches it locally; subsequent queries on the same passages are sub-20ms warm.

Build details

Field	Value
Encoder	Alibaba-NLP/gte-modernbert-base 768d, CLS-pooled, L2-normalised
Encoder revision	e7f32e3c00f91d699e8c43b53106206bcc72bb22 (pinned)
Format	rlat knowledge-model v4 (ZIP + JSON + NPZ)
Storage mode	remote (source pinned at SHA, fetched on demand, SHA-verified)
Source repo	MicrosoftDocs/PowerShell-Docs
Source scope	reference
Source commit	58137d73023e752c811cd327a91e03065844dc75
Source branch	main (commit SHA-pinned; reproducible regardless of branch movement)
Files indexed	2,647
Passages	107,033
Build date	2026-04-28
Built on	Kaggle T4 (GPU encoding, batch_size=64, runtime=torch)
File size	639.7 MB

Usage

Single-hop search

rlat search powershell-docs.rlat "what does X do?" --top-k 5

Skill-context (Anthropic skill !command block)

!`rlat skill-context powershell-docs.rlat --query "$user_query" --top-k 5`

The output is markdown with citation anchors, drift status, and ConfidenceMetrics — ready for an LLM to ground on.

Multi-hop deep-search

rlat deep-search powershell-docs.rlat "harder cross-file question" --max-hops 3

Requires an Anthropic API key. See the deep-search docs.

Refreshing against upstream

This model pins to the source commit 58137d73. To re-index against the current upstream tip:

# Option A: rebuild on Kaggle's free T4 (recommended for big corpora)
# See the rlat-build-on-kaggle skill at:
# https://github.com/tenfingerseddy/resonance-lattice/tree/main/.claude/skills/rlat-build-on-kaggle

# Option B: rebuild locally
pip install rlat[build,ann]
rlat install-encoder
git clone --depth 1 -b main https://github.com/MicrosoftDocs/PowerShell-Docs.git src/
rlat build src/reference -o powershell-docs.rlat \
  --store-mode remote \
  --remote-url-base https://raw.githubusercontent.com/MicrosoftDocs/PowerShell-Docs/<NEW_SHA>/reference/ \
  --runtime torch

Honest limits

• The encoder is gte-modernbert-base 768d with no per-corpus optimisation. Default retrieval is dense cosine over the base band — single recipe, no rerankers, no lexical sidecar.
• For per-corpus retrieval lift, you can run rlat optimise locally to add a 512d MRL-trained band on top of this archive (opt-in, costs API + GPU time). See docs/user/OPTIMISE.md.
• Drift detection is automatic: if the source files at GitHub change, query results show a drifted status until the model is rebuilt against the new commit.

License

The rlat software is licensed under BSL 1.1 (Business Source License — source-available; production use of the licensed work is permitted up to the parameters in LICENSE.md).

This .rlat archive contains embeddings + metadata + a SHA-pinned URL manifest; source bytes are NOT bundled and are fetched from upstream GitHub at query time, where the upstream repository's license applies.