ente-ai/ensu-knowledge-packs

Ensu Knowledge Packs

Prebuilt on-device retrieval indexes ("knowledge packs") for Ensu, ente's private on-device AI assistant — plus the scripts that generate them. Ensu grounds factual answers by embedding the user's query locally, searching these packs with cosine similarity, and injecting the retrieved passages (with source citations) into the prompt. Everything runs on-device; no query ever leaves the phone.

Layout

Each dataset lives in its own self-contained folder: generation scripts, the generated pack, and dataset-specific docs together. Code shared across packs (the embedder, index writer, dump utilities) lives in shared/.

Folder	Source	Granularity	Rows	Pack size
fullwiki/	Full English Wikipedia, 2026-10-01 snapshot	lead-only (one passage per article)	528,303	347.2 MiB
wikibooks/	English Wikibooks, 2026-10-01 snapshot	section-chunk (curated, ≤1,400 chars, incl. Cookbook)	250,000	193.2 MiB
simplewiki/	Simple English Wikipedia, 2026-10-01 snapshot	lead-only (one passage per article)	265,278	161.6 MiB

fullwiki answers what (definitions, facts) across the head and the mid-tail the on-device model doesn't know; wikibooks answers how/why (explanations, worked examples, procedures). They share model, dim, and quantization, so their cosine scores are comparable and results merge into one global top-k.

fullwiki is the larger general-reference pack, with broader topic coverage in full-English prose. Its selection includes a Simple-English safety net. simplewiki remains the smaller, simpler-language option.

Inside a dataset folder:

• scripts/ — build.py (dump → rows → embeddings → pack) and search.py (quality eval / threshold tuning), plus any dataset-specific analysis tools
• data/ — the generated pack: vectors.i8, meta.zst, meta.offsets, manifest.json

Shared: shared/ — common.py (embedder, fresh batched embedding, block-zstd meta writer, MetaReader, dump helpers) and requirements.txt.

Refresh all three packs

Every refresh downloads the latest Wikimedia snapshots again, verifies the published SHA-1 checksums, and rebuilds rankings, passages, and embeddings from scratch. There is no resume or cache reuse. An interrupted run must restart in a new output directory. The GGUF model is a fixed build dependency, not a source snapshot.

Large static downloads use short, version-checked requests when supported by the server. Every range is fetched anew in each invocation; no download progress is restored from disk. An interrupted request is discarded and requested again in full, and the completed XML snapshot must match Wikimedia's checksum.

python scripts/refresh.py --output /tmp/ensu-refresh-YYYYMMDD \
  --gguf /path/to/embeddinggemma-300M-Q8_0.gguf \
  --months 202607 202608 202609 \
  --questions /path/to/evaluation-questions.json

Choose three completed pageview months available upstream and supply your local evaluation questions (JSON objects with q and category fields). The output directory must not exist. The runner builds Simplewiki and Wikibooks, then selects and builds Fullwiki using the fresh Simplewiki safety net. It validates every pack, runs retrieval for the evaluation questions, and removes temporary downloads and ranking files after success. It keeps the staged packs, logs, source provenance, checksums, and evaluation results. Review those results before replacing the published data/ artifacts; publication is a separate step.

Individual builders also always download and parse afresh. They remove the source download and legacy parse/embedding caches after successful completion.

For slow source connections, the runner accepts --dump-mirror (a mirror root containing the per-wiki directories) and --pageviews-base-url (its monthly pageview directory). Choose a Wikimedia-listed mirror. Latest snapshot dates and XML checksums still come directly from Wikimedia; the actual download URLs are recorded for audit.

The snapshot date in each manifest describes source freshness; a refresh does not turn these offline packs into a live feed. The manifest also records the source URL and checksum, build time, model checksum, and build dependencies.

Ensu pins pack downloads to an immutable Hugging Face revision. After publishing a validated refresh, update its knowledge catalog with that revision, each pack's new dataset identity, and artifact sizes and SHA-256 hashes. The app's artifact order is manifest.json, vectors.i8, meta.zst, meta.offsets. Publishing new files on the dataset's main branch alone does not update the app's download catalog.

October 2026 refresh checks

All three packs passed metadata/vector alignment, dimension, norm, count, and artifact checksum checks. The refresh and shared-code tests passed (16 tests). The same 982 evaluation questions were searched against each refreshed pack. Of 2,946 top results, 2,620 retained the exact query and full passage from the July evaluation; their relevance labels were retained. The other 326 results received new LLM relevance judgments after reading the complete passages.

Pack	Gate	Precision	Recall	F1	Noise above gate (of 210)
Simplewiki	0.55	0.7899	0.8805	0.8328	14
Wikibooks	0.58	0.8588	0.6995	0.7710	22
Fullwiki	0.55	0.8091	0.8990	0.8517	10

The existing gates are retained. Simplewiki and Wikibooks were essentially stable; Wikibooks had one additional noise query above its gate. Fullwiki's F1 declined modestly from 0.8579 to 0.8517. Of 14 results that lost relevance, 11 previously retrieved titles were absent from the refreshed corpus and three remained but were outranked. Its fresh pageview selection added 44,358 titles and removed 42,210.

This is a retrieval regression review, not an independent full recalibration or a factual audit of every source passage. Metrics use stored cosine scores rounded to four decimals, matching the July evaluation.

Pack format

Exactly what the Ensu Rust RetrievalIndex reads:

• vectors.i8 — raw row-major count × dim int8; L2-unit vectors × 127
• meta.zst — row metadata as independent zstd frames ("blocks") of meta_rows_per_block JSONL rows {id, title, text} each, row-aligned to vectors. Row lookup: block = row / meta_rows_per_block; decompress the frame at offsets[block]..offsets[block+1]; take line row % meta_rows_per_block. A top-k query decompresses at most k blocks (~0.2 ms each on-device).
• meta.offsets — little-endian u64 array of n_blocks + 1 byte offsets into meta.zst (last entry = file size)
• manifest.json — dim, count, quant, scale, meta_codec, meta_rows_per_block, url_template, dataset (with dump date), granularity, model + prompt provenance

Rows carry no url: it is reconstructed as url_template.format(title=percent-encoded title, spaces → underscores). Packs may add fields the reader ignores if unused — e.g. wikibooks rows carry a section, appended to the reconstructed URL as a # anchor.

Embedding model

Documents are embedded with EmbeddingGemma-300M Q8_0 (ggml-org/embeddinggemma-300M-GGUF) via llama.cpp — the same model artifact Ensu loads on-device for query embedding, so document and query vectors come from a bit-identical model.

• Document prompt: title: {title} | text: {text}
• Query prompt: task: search result | query: {query}
• Mean pooling, L2 normalization
• Flash attention enabled for embedding; recorded in the manifest
• Matryoshka truncation to 512 dims: keep the first 512 of 768 components, re-normalize. The query side must do the same (truncating an already-normalized 768-dim vector and re-normalizing is equivalent).

Both prompts and the truncation are recorded in each pack's manifest.json.

License

• Scripts: MIT
• Wikipedia & Wikibooks content: CC BY-SA 4.0 (attribution is satisfied in-app by citing each passage's source URL)