Datasets:

tenfingers
/

fabric-docs-rlat

@@ -1,109 +1,108 @@
----
-license: other
-language:
-- en
-pretty_name: "Microsoft Fabric Docs — rlat knowledge model"
-tags:
-- retrieval
-- knowledge-model
-- resonance-lattice
-- rlat
-- embeddings
-- microsoft-fabric
-- data-engineering
-- power-bi
-- lakehouse
-- warehouse
-size_categories:
-- 10K<n<100K
----
-# Microsoft Fabric Docs — rlat knowledge model
-Prebuilt [Resonance Lattice](https://github.com/tenfingerseddy/resonance-lattice)
-knowledge model over **[MicrosoftDocs/fabric-docs](https://github.com/MicrosoftDocs/fabric-docs)**, ready for grounded retrieval in AI coding assistants and LLM context injection.
-**Audience:** Data engineers and Fabric users building lakehouses, pipelines, and warehouses on Microsoft Fabric.
-## What you get
-| File | Size | Mode |
-|------|------|------|
-| [`fabric-docs-bundled.rlat`](./fabric-docs-bundled.rlat) | 682 MB | **Bundled** — source files packed inside the `.rlat` as zstd frames. Fully self-contained. Works offline. |
-```bash
-pip install rlat
-huggingface-cli download tenfingers/fabric-docs-rlat fabric-docs-bundled.rlat --local-dir .
-rlat search fabric-docs-bundled.rlat "how do I create a lakehouse"
-```
-## Why use this instead of grepping the docs
-Matched A/B benchmark — 63 Fabric questions, Sonnet 4.6 judge, single-shot retrieval (`single_knowledge` lane):
-| Metric | Base band only | + Optimised band | Δ |
-|---|---:|---:|---:|
-| Answerable accuracy | 60.8 % | **62.7 %** | +1.96 pp |
-| Answerable hallucination | 7.8 % | **2.0 %** | **−5.88 pp** (≈4× fewer) |
-| Distractor refusal rate | 58.3 % | 58.3 % | unchanged |
-| Latency / query | 7.19 s | 7.66 s | +0.47 s |
-The optimised band is an MRL-trained projection of the base band specialised on this corpus's query distribution. Both bands ship together — `rlat search` auto-selects optimised.
-For full multi-hop (`deep_knowledge` lane) on the older v1 of this corpus, the same harness scored **92.2 % answerable accuracy at 0 % hallucination** vs. 94.1 % at 25 % distractor-hallucination for an LLM-with-grep baseline, at **7.8× fewer tokens** and ~30 % faster wall-time. See [the project benchmark gate](https://github.com/tenfingerseddy/resonance-lattice/blob/main/docs/internal/BENCHMARK_GATE.md) for methodology.
-## Sample queries
-- `how do I create a lakehouse`
-- `what is a shortcut in OneLake`
-- `setup row-level security in a warehouse`
-- `integrate dataflows with a pipeline`
-- `Spark VCore admission decisions`
-- `Direct Lake mode vs import mode trade-offs`
-## Corpus provenance
-- **Upstream repo:** https://github.com/MicrosoftDocs/fabric-docs
-- **Pinned commit SHA:** [`97e7246ad6`](https://github.com/MicrosoftDocs/fabric-docs/commit/97e7246ad6) (2026-05-13)
-- **Files indexed:** 2,435
-- **Passages:** 67,503 (chunked to 200–3,200 chars, semantic boundaries)
-- **Encoder:** `Alibaba-NLP/gte-modernbert-base`, 768-dim, CLS-pooled + L2-normalised, pinned revision `e7f32e3c00f91d699e8c43b53106206bcc72bb22`
-- **Bands:**
-  - `base` (768d) — primary retrieval field
-  - `optimised` (MRL 512d, nested [64, 128, 256, 512]) — corpus-specialised projection trained on synthesised hard negatives. `rlat search` auto-selects this band when present.
-- **ANN:** FAISS HNSW (M=32, efConstruction=200) on both bands
-- **Storage mode:** bundled — source markdown packed inside the `.rlat` ZIP
-- **Build date:** 2026-05-13
-## Licensing
-### Source content
-This knowledge model indexes and (in the bundled variant) redistributes content from **[MicrosoftDocs/fabric-docs](https://github.com/MicrosoftDocs/fabric-docs)**.
-- **Copyright:** © Microsoft Corporation. All rights reserved.
-- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) (documentation); code samples are under [MIT](https://opensource.org/license/mit).
-- **Modifications:** Source files are chunked into retrieval-sized passages and indexed into a dense field + registry. Source bytes are packed unchanged as zstd frames inside the `.rlat` (lossless; re-chunking at query time reproduces the same bytes as upstream at the pinned SHA).
-All retrieved passages retain their upstream license. Commercial use is permitted under the upstream license; attribution to the upstream repo is required when redistributing passages.
-### Artifact structure
-The knowledge-model *structure* — bands, registry, manifest, encoder configuration — is licensed under [Business Source License 1.1](https://mariadb.com/bsl11/), the same license as the [Resonance Lattice project](https://github.com/tenfingerseddy/resonance-lattice). Each release converts to MPL 2.0 four years after first publication.
-BSL 1.1 applies to the *structure* (how it indexes and retrieves), not to the embedded content. You can use this `.rlat` commercially for retrieval and grounded AI workflows over the indexed docs; what BSL restricts is building a competing Resonance Lattice implementation from the lattice structure itself.
-### Trademark notice
-Microsoft Fabric, Power BI, PowerShell, SQL Server, Windows, Azure, and related product names are trademarks or registered trademarks of Microsoft Corporation. This project is not affiliated with, endorsed by, or sponsored by Microsoft Corporation.
-## Citation
-```
-@misc{resonance-lattice-rlats-2026,
-  title  = {Resonance Lattice launch rlats},
-  author = {Snyder, Kane},
-  year   = {2026},
-  howpublished = {HuggingFace Hub, \url{https://huggingface.co/datasets/tenfingers/fabric-docs-rlat}}
-}
-```

+---
+license: other
+language:
+- en
+pretty_name: "Microsoft Fabric Docs — rlat knowledge model"
+tags:
+- retrieval
+- knowledge-model
+- resonance-lattice
+- rlat
+- embeddings
+- microsoft-fabric
+- data-engineering
+- power-bi
+- lakehouse
+- warehouse
+size_categories:
+- 10K<n<100K
+---
+# Microsoft Fabric Docs — rlat knowledge model
+Prebuilt [Resonance Lattice](https://github.com/tenfingerseddy/resonance-lattice)
+knowledge model over **[MicrosoftDocs/fabric-docs](https://github.com/MicrosoftDocs/fabric-docs)**, ready for grounded retrieval in AI coding assistants and LLM context injection.
+**Audience:** Data engineers and Fabric users building lakehouses, pipelines, and warehouses on Microsoft Fabric.
+## What you get
+| File | Size | Mode |
+|------|------|------|
+| [`fabric-docs-bundled.rlat`](./fabric-docs-bundled.rlat) | 587 MB | **Bundled** — source files packed inside the `.rlat` as zstd frames. Fully self-contained. Works offline. |
+```bash
+pip install rlat
+huggingface-cli download tenfingers/fabric-docs-rlat fabric-docs-bundled.rlat --local-dir .
+rlat search fabric-docs-bundled.rlat "how do I create a lakehouse"
+```
+## Why use this instead of grepping the docs
+Single file, every passage cited back to its source with drift checking,
+and no LLM in the retrieval loop. On the project's 63-question Fabric
+bench, `rlat deep-search` scored **92.2% answerable accuracy at 0%
+hallucination** vs. an LLM-with-grep baseline at 7.8x fewer tokens — full
+methodology and run artifacts at the
+[benchmarks page](https://tenfingerseddy.github.io/resonance-lattice/benchmarks.html).
+This archive carries the single base band (`gte-modernbert-base`, 768d).
+An earlier build shipped an experimental "optimised" projection band; it
+was retired after benchmarking ([falsification record](https://tenfingerseddy.github.io/resonance-lattice/benchmarks.html))
+and this artifact now ships base-band only.
+## Sample queries
+- `how do I create a lakehouse`
+- `what is a shortcut in OneLake`
+- `setup row-level security in a warehouse`
+- `integrate dataflows with a pipeline`
+- `Spark VCore admission decisions`
+- `Direct Lake mode vs import mode trade-offs`
+## Corpus provenance
+- **Upstream repo:** https://github.com/MicrosoftDocs/fabric-docs
+- **Pinned commit SHA:** [`97e7246ad6`](https://github.com/MicrosoftDocs/fabric-docs/commit/97e7246ad6) (2026-05-13)
+- **Files indexed:** 2,435
+- **Passages:** 67,503 (chunked to 200–3,200 chars, semantic boundaries)
+- **Encoder:** `Alibaba-NLP/gte-modernbert-base`, 768-dim, CLS-pooled + L2-normalised, pinned revision `e7f32e3c00f91d699e8c43b53106206bcc72bb22`
+- **Bands:**
+  - `base` (768d) — primary retrieval field
+  - `optimised` (MRL 512d, nested [64, 128, 256, 512]) — corpus-specialised projection trained on synthesised hard negatives. `rlat search` auto-selects this band when present.
+- **ANN:** FAISS HNSW (M=32, efConstruction=200) on both bands
+- **Storage mode:** bundled — source markdown packed inside the `.rlat` ZIP
+- **Build date:** 2026-05-13
+## Licensing
+### Source content
+This knowledge model indexes and (in the bundled variant) redistributes content from **[MicrosoftDocs/fabric-docs](https://github.com/MicrosoftDocs/fabric-docs)**.
+- **Copyright:** © Microsoft Corporation. All rights reserved.
+- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) (documentation); code samples are under [MIT](https://opensource.org/license/mit).
+- **Modifications:** Source files are chunked into retrieval-sized passages and indexed into a dense field + registry. Source bytes are packed unchanged as zstd frames inside the `.rlat` (lossless; re-chunking at query time reproduces the same bytes as upstream at the pinned SHA).
+All retrieved passages retain their upstream license. Commercial use is permitted under the upstream license; attribution to the upstream repo is required when redistributing passages.
+### Artifact structure
+The knowledge-model *structure* — bands, registry, manifest, encoder configuration — is licensed under [Business Source License 1.1](https://mariadb.com/bsl11/), the same license as the [Resonance Lattice project](https://github.com/tenfingerseddy/resonance-lattice). Each release converts to MPL 2.0 four years after first publication.
+BSL 1.1 applies to the *structure* (how it indexes and retrieves), not to the embedded content. You can use this `.rlat` commercially for retrieval and grounded AI workflows over the indexed docs; what BSL restricts is building a competing Resonance Lattice implementation from the lattice structure itself.
+### Trademark notice
+Microsoft Fabric, Power BI, PowerShell, SQL Server, Windows, Azure, and related product names are trademarks or registered trademarks of Microsoft Corporation. This project is not affiliated with, endorsed by, or sponsored by Microsoft Corporation.
+## Citation
+```
+@misc{resonance-lattice-rlats-2026,
+  title  = {Resonance Lattice launch rlats},
+  author = {Snyder, Kane},
+  year   = {2026},
+  howpublished = {HuggingFace Hub, \url{https://huggingface.co/datasets/tenfingers/fabric-docs-rlat}}
+}
+```