2.4.0 RAG folder picker, Ollama and vLLM, Bedrock sampling

The document folder of a RAG setup is chosen in a folder picker, Claude on AWS Bedrock gets one sampling option per request, the RAG runtime stores no host name in the text of a failure, and Ollama and vLLM definitions take their server from the container only - a reasoning model served by Ollama answers.

Upgrading from 2.3.2

Only where it applies to your site; nothing else changes.

1. Update the checkout to 2.4.0 and reinstall the CLI: pip install -e "Model-Definition-Builder/cli[viya]".
2. Where claude_haiku_4_5_bedrock is deployed: mdb register claude_haiku_4_5_bedrock --update and mdb publish claude_haiku_4_5_bedrock --wait, then restart its Deployment.
3. Where RAG is used: mdb builders-import (or mdb content-deploy --only RAG --to "<your RAG content root>") brings the changed rag_core, ingestion job and Prompt Builder; then reload the Prompt Builder report. Rows written before keep their text, and a setup keeps the folder it was saved with.
4. For Ollama and vLLM definitions of your own (the release ships none), per definition: mdb generate <id>, mdb register <id> --update, mdb publish <id> --wait, then mdb deploy <id> with OLLAMA_BASE_URL / VLLM_BASE_URL set to the server the cluster reaches, and apply the YAML. The base_url option leaves options.json and the Prompt Builder. For a reasoning model (gpt-oss, Qwen3 and the like), add it again with mdb add first - or set reasoning_effort, max_tokens: 4000 and runtime.timeout_s: 300 by hand.
5. Optional, where RAG is used: set RAG: start folder on the compute server and RAG: start folder in SAS Content in the Options pane of the Prompt Builder object, and tick the allow folders above options where users may go above them (The folder picker and its compute session).

Added

• A folder picker for the document folder of a RAG setup. Browse lists the folders the user may read, of the compute server or of SAS Content; the folder can no longer be typed. The compute server is listed in a session of the ingestion compute context, started as soon as the RAG kind is shown (it takes about half a minute), pinged every minute, deleted when the page shows a template or is left, and ended by SAS Viya after five idle minutes.
• Four options for where the picker starts: ragServerFolder, ragServerBrowseAbove, ragContentFolder, ragContentBrowseAbove. A start folder is blank by default; set, with its allow-above option off, the picker stays below it and a setup can only be saved with a folder below it.

Changed

• The server of an Ollama or vLLM definition is no scoring option. The score code reads OLLAMA_BASE_URL / VLLM_BASE_URL, else the default of the definition; a base_url a caller sends is ignored. The Prompt Builder no longer shows the server's address as a field and no longer sends it with every call.
• mdb deploy writes the server into the Deployment of an Ollama or vLLM definition, and the token (OLLAMA_API_KEY / VLLM_API_KEY) as an optional reference to the key Secret.
• mdb package-check: an option label the shipped Prompt Builder report lacks is drift, as a changed file is: the report gains it when it is saved in Visual Analytics with the newer page before the export of a release. --release still fails for it, and mdb package-export still refuses such a report.

Fixed

• Bedrock no longer sends temperature and topP together. Claude refuses a request with both, as on Anthropic's own API (fixed there in 1.1.0). Both Bedrock templates send the one the caller set, temperature where the caller set both or neither; claude_haiku_4_5_bedrock is regenerated.
• The RAG runtime keeps host names out of the text of a failure. A database or HTTP error names the server it failed at - SingleStore the internal host and port of a node. error_text of the ledger and the step tables, rag_runs, rag_doc_events and the message of the ingestion job now hold <host> in its place; file names in the text stay.
• The Logging README says what the collection-script tests need (bash and GNU sed) and that they are skipped without them.
• A reasoning model added from Ollama answers. mdb add ollama reads /api/show: a model that thinks gets reasoning_effort, max_tokens 4000 and 300 seconds per call (gpt-oss:20b answered nothing with 1000 tokens and timed out at 60 seconds with 4000), an embedding model its vector and context length. vLLM definitions take the context length from /models.
• An answer cut off by the output budget fails, naming the remedy, instead of returning an empty response.
• A server that refuses the token names the variable to set (OLLAMA_API_KEY / VLLM_API_KEY) instead of Request failed: {responseObject.status_code}, and mdb deploy prints the command that creates the key Secret, not only the patch of an existing one.
• mdb validate --live refuses an Embedding_Length the server does not return (a vLLM embedding definition kept the 768 fallback).
• The model card of an Ollama or vLLM definition named an endpoint_url option that never existed.