.8. 8 8888 `8.`8888. ,8' d888888o.
.888. 8 8888 `8.`8888. ,8'.`8888:' `88.
:88888. 8 8888 `8.`8888. ,8' 8.`8888. Y8
. `88888. 8 8888 `8.`8888.,8' `8.`8888.
.8. `88888. 8 8888 `8.`88888' `8.`8888.
.8`8. `88888. 8 8888 `8. 8888 `8.`8888.
.8' `8. `88888. 8 8888 `8 8888 `8.`8888.
.8' `8. `88888. 8 8888 8 8888 8b `8.`8888.
.888888888. `88888. 8 8888 8 8888 `8b. ;8.`8888
.8' `8. `88888. 8 888888888888 8 8888 `Y8888P ,88P' > local runtime / grounded exports
autonomousAI data prep.
Alys runs from the terminal and turns messy knowledge into grounded, AI-ready data.
Autonomous infrastructure that ingests files, raw documents, structured XML, and audio/video transcripts—deduplicating, canonicalizing, and exporting clean corpora for RAG, fine-tuning, embeddings, and evaluation.
Install Alys
> npx alys-akusa prepare ./company-docsLive Prepared Run
Autonomous data preparation infrastructure, run directly from the terminal.
The web dashboard handles authentication, usage, observability, and job metadata, while the terminal remains the runtime. Local preparation writes artifacts under ~/Alys/prepared/{run_id} and does not spend hosted generation credits.
output workspace
~/Alys/prepared/56b9a700be3d
> npx --yes alys-akusa@latest prepare . --yes --workspace ~/Alys
[VALIDATE]0 Git LFS pointer files remaining
[PAIR]189 audio files with 100% transcript coverage
[EXPORT]122,376 records across fine-tune, RAG, QA, media, and review artifacts
[READY]0 critical media diagnostics
990
files indexed
materialized checkout
801
documents parsed
docs + transcripts
9.66M
tokens parsed
full corpus pass
22,306
chunks created
raw chunk graph
20,396
canonical chunks
deduped accepted
122,376
export records
all profiles written
Media Readiness
189
audio files
MP3 corpus ready
100%
transcript coverage
.nlp sidecars paired
77/100
media readiness
with no criticals
0
critical diagnostics
blocking issues cleared
Dataset Correction
Git LFS pointers
0 remaining
real media files, not placeholders
earnings21/media
44 MP3s
materialized and indexed
earnings22/media
125 MP3s
valid MP3 headers
subset10/media
10 symlinks
points into earnings22/media
Transcript sidecars
179 symlinks
audio paired with existing .nlp files
Export Stack
openai-finetune.jsonlanthropic-instruction.jsonlrag-chunks.jsonlqa-dataset.jsonlevaluation-set.jsonlmedia-manifest.jsontranscription-plan.jsonreadiness-report.jsonretrieval-simulation.jsonquality-benchmarks.jsonKnown Warnings
431
Non-blocking warnings remain: missing media metadata sidecars, untimestamped transcript sidecars, and unavailable ffprobe. The important part: no critical media diagnostics.
Pipeline
Everything your model needs before retrieval, training, or evaluation.
>INGEST
Files, folders, raw documents, structured XML, and transcript-backed audio/video assets
terminal runtime
>PROCESS
Chunk massive corpora, deduplicate content, canonicalize records, and evaluate AI readiness
grounded engine
>EXPORT
Corpora for RAG, evaluation, OpenAI/Anthropic fine-tuning JSONL, embeddings, and quality benchmarks
AI-ready
Local Runtime
user@machine repo % npx alys-akusa prepare ./knowledge-base
[INGEST]ingest files/folders to prepare embeddings and eval profiles
[AUDIT]audit quality, identify duplicate content, and find grounding gaps
[TRANSCRIBE]build Google STT V2 transcription plans and media manifests
[SIMULATE]simulate retrieval behavior for RAG evaluation at scale
[IMPROVE]improve corpus content, canonicalize records, and clean text
[BENCHMARK]benchmark quality indicators for original vs. improved data
[READY]exports written with source fingerprints
Key Metrics
98
files indexed
436k
tokens parsed
957
chunks accepted
88%
grounding score
Get Started