feat: implement local inference mode with SAP RPT-1 OSS model

[KoblerS]

Add Local Inference Mode with SAP RPT-1 OSS Model

New Feature

✨ Introduces a local inference mode for the CAP AI plugin, allowing developers to run predictions using the open-source SAP RPT-1 OSS model without requiring a full SAP AI Core connection. Two backends are supported: a local Python subprocess (AICore-local) and the HuggingFace Inference API (AICore-hf).

Changes

• srv/LocalRPTService.js: New service implementation with two backend classes:

  • LocalSubprocessRPTService — spawns a Python subprocess (infer.py), automatically downloads and caches the model checkpoint (~65 MB) and sentence embedder on first startup, and communicates via a JSON stdin/stdout protocol.
  • HFInferenceRPTService — calls the HuggingFace Inference API directly, no local Python required.
  • LocalRPTService (default export) — delegates to either backend based on the local: true config flag.

• lib/rpt/infer.py: New Python inference server script that reads JSON requests from stdin, runs predictions using SAP_RPT_OSS_Classifier, and writes JSON responses to stdout. Includes dtype coercion, top-3 candidate ranking, and a readiness signal on startup.

• srv/LocalRPTService.cds: New CDS definition wiring the AICore service to the LocalRPTService.js implementation.

• package.json: Added AICore-local and AICore-hf CDS service kinds. Changed the default (no-profile) kind from AICore-mocked to AICore-local. Production and hybrid profiles continue to use AICore-btp.

• README.md: Added a "Local mode (without SAP AI Core)" section documenting both backends, setup steps (licence acceptance, HF token creation, .cdsrc-private.json configuration), and installation instructions for the Python package.

• .gitignore: Fixed an overly broad resources/ ignore rule to only ignore resources/.cds-private.json.

• 🔄 Regenerate and Update Summary

PR Bot Information

Version: 1.23.0

• Summary Prompt: Default Prompt
• Output Template: Default Template
• LLM: anthropic--claude-4.6-sonnet
• File Content Strategy: Full file content
• Correlation ID: 70c2036e-f7c3-4036-857c-28f2e4a906e8
• Event Trigger: pull_request.opened

[KoblerS]

TODO: add tests

[hyperspace-pr-bot]

The PR introduces a useful local-inference mode but has several correctness issues that need addressing before merge: a typo in .gitignore that will leave sensitive files unignored, a missing module-level pandas import in the Python script, a null-dereference race on the subprocess stdin, a token-resolution inconsistency between _hfToken() and the documented credential path, and — most impactful — changing the plugin's default AICore kind from the safe AICore-mocked to AICore-local, which will silently trigger model downloads and Python subprocess spawning in all downstream projects that don't override the kind.

PR Bot Information

Version: 1.23.0

• File Content Strategy: Full file content
• LLM: anthropic--claude-4.6-sonnet
• Correlation ID: 70c2036e-f7c3-4036-857c-28f2e4a906e8
• Event Trigger: pull_request.opened

[KoblerS]

TODO: add huggingface secret for read access to rpt1 oss model

[KoblerS]

Unfortunately the intereference Python script is required, since the model has been only published as PyTorch model and not an ONNX model.