Documentation
Build, inspect, resume, evaluate, and export chat and embedding workflows.
Start with deterministic offline tutorials, then move through concepts, task guides, generated references, operations, and security. Paid, network, model-download, GPU, upload, remote-code, and future remote-execution paths stay explicitly gated.
Start
Concepts
Core concepts
Understand records, codecs, templates, manifests, provenance, runs, and artifacts shared by both tracks.
Open guideFull tool trajectories
Generate and validate the canonical tool-calling dataset shape, including tool calls, tool results, and final assistant responses.
Open guideEmbedding recipes and data shapes
Choose losses, prompts, pooling, dimensions, mining, and model-specific boundaries honestly.
Open guideModel recipe qualification
Configure, authorize, smoke, qualify, and support model recipes without overstating evidence.
Open guideTutorials
Chat response distillation and SFT
Run a complete offline chat workflow, then understand the boundary for provider teachers and production recipes.
Open guide10-minute offline embeddings
Validate retrieval pairs, train and resume the CPU fixture, compare evaluation, and verify export.
Open guideExamples
Use the bundled receptionist and retail support scenarios, plus the provider config example, as starting points for your own datasets.
Open guideHow-to
Validation
Understand what the dataset validator checks, what summary output means, and how validation fits into generation and translation workflows.
Open guideChoose a distillation target
Select response, vector, score, or ranking distillation and preserve the controls each target needs.
Open guideTranslation
Translate generated datasets with the offline local-pseudo path or explicit provider-backed OpenAI and Anthropic adapters.
Open guideEvaluate runs and verify artifacts
Compare baselines, interpret task metrics, classify checkpoints, and verify release evidence.
Open guideProvider config
Configure OpenAI and Anthropic persona generation, simulation, and translation through config files or explicit command flags.
Open guideScenario authoring
Define assistant behavior through scenario JSON, including personas, tool schemas, goals, and stopping rules.
Open guideReference
Complete embedding CLI command reference
Exact syntax and effects for every registered embedding command.
Open guideCLI reference
Review the supported commands, their high-level purpose, and the most important flags for each workflow.
Open guideTypeScript SDK and API
Use the stable packed exports and understand browser-safe versus Node-only boundaries.
Open guideArchitecture
Understand canonical data, distillation, orchestration, Python training, evaluation, and remote-execution boundaries.
Open guideCompatibility and reproducibility
Understand package versions, schemas, protocols, configuration precedence, errors, and benchmark limits.
Open guidePython trainer and API
Install the separate wheel, validate versioned jobs, and run or verify local artifacts.
Open guideConfiguration and schemas
Resolve configuration safely and locate every packaged versioned contract.
Open guideModels, recipes, providers, and execution
Read exact lock evidence and distinguish available frameworks from unavailable production recipes.
Open guideRunPod execution and support status
Distinguish the available offline and read-only RunPod contracts from unavailable live execution.
Open guideOperations
Security
Log conversion deferred
Understand why real-log conversion is intentionally unavailable and what must exist before it can be released safely.
Open guideSecurity, privacy, and compliance
Threat-model secrets, sensitive data, provider output, artifacts, retention, deletion, and legal obligations.
Open guideProject
Docs site maintenance
Run, edit, validate, and build the embedded Astro documentation site that now lives inside the repository root.
Open guideMigration, support, and release notes
Preserve legacy workflows, assess compatibility, and rehearse a private-alpha release safely.
Open guideGlossary
Use consistent terms across chat, embeddings, manifests, training, and release operations.
Open guideDocumentation acceptance checklist
Reproduce clean-package user tasks and responsive production-site behavior.
Open guide