Skip to content
Nibiru docsv0.9.2
ENDEJAESFR

AI Roadmap

Where Nibiru's AI integration is going — the plan that gets us from RAG-grounded Oracle to a fine-tuned LoRA in production.

Stable Reading time ~ 2 min Edit on GitHub

Nibiru’s ambition: be the first PHP framework with a fine-tuned model trained on its own knowledge, served as a first-class part of the dev experience. This page tracks the steps.

Phase 1 — Today: RAG Oracle ✓

Section titled “Phase 1 — Today: RAG Oracle ✓”
  • Markdown chunker with H2/H3 boundaries.
  • OpenAI embeddings (text-embedding-3-small).
  • Vector index as a single JSON file.
  • Astro endpoint /api/oracle calling Claude with retrieved context.
  • Floating chat widget on every doc page.
  • Multilingual (EN/DE/JA/ES/FR) input + output.

Why first. RAG works without training, scales linearly with content size, and is dirt cheap. Every doc edit improves answer quality the same hour.

Phase 2 — Next: Public corpus + LoRA recipe

Section titled “Phase 2 — Next: Public corpus + LoRA recipe”
  • npm run build:corpus ships in main (instructions/chat/chunks JSONL).
  • Published Hugging Face dataset (nibiru-framework/docs-corpus).
  • Reference Axolotl YAML for Llama 3.1 8B.
  • Reference recipe for Qwen 2.5 7B and Mistral Nemo 12B.
  • Eval set: 200 hand-graded Nibiru questions with golden answers.

Why second. Once the corpus is reproducible from docs, anyone can train. We treat docs as the source of truth and the corpus as a derivative artifact.

Phase 3 — Then: Hosted LoRA endpoint

Section titled “Phase 3 — Then: Hosted LoRA endpoint”
  • Train a first-pass LoRA on the public corpus.
  • Serve via vLLM behind /api/oracle with a feature flag.
  • Side-by-side UI comparing Claude (RAG) vs LoRA (no RAG) vs LoRA + RAG.
  • Telemetry: which form does the user prefer per question type?

Why third. Side-by-side comparison reveals where the LoRA helps (idiomatic Nibiru style) and where it hurts (very long context, fresh edits not yet retrained).

Phase 4 — Eventually: Editor agents

Section titled “Phase 4 — Eventually: Editor agents”
  • PHPStorm plugin: highlight a controller, ask the Oracle to convert it to a module.
  • CLI agent: ./nibiru ask "rewrite this controller as a JSON endpoint".
  • PR review bot: explain Nibiru-specific deviations in pull requests on framework forks.

Phase 5 — Aspirational: Active learning

Section titled “Phase 5 — Aspirational: Active learning”
  • User feedback in the chat widget (👍 / 👎) writes a row to a private dataset.
  • Weekly review queue surfaces low-rated answers for human annotation.
  • Improved answers re-enter the corpus on the next training cycle.

How to help

Section titled “How to help”
  • Ask the Oracle hard questions and rate the answers.
  • Open issues on the GitHub repo for missing topics.
  • Contribute translations — every translated doc page is also a parallel corpus row.
  • Try a LoRA fine-tune on the published corpus and share results.