refactor: reduce configuration sprawl (dead run-mode/chunking config + unify LanceDB table name)

[jdye64]

Description

Incremental reduction of configuration sprawl in the repo. Grouped into logical commits; each addresses a specific finding from the configuration audit.

1. Remove dead ModelRunMode run-mode vocabulary

ModelRunMode = "local" | "NIM" | "build-endpoint" was dead metadata: every BaseModel subclass hard-coded model_runmode → "local", and .model_runmode was only ever read in tests. Real local/NIM/hosted selection happens per-stage in operators/operator_archetype.py via *_invoke_url presence + GPU detection. Removed the type alias, the abstract property on BaseModel, and its 10 concrete overrides; tests updated. This removes one of the four competing "run mode" vocabularies. (Report suggestion #2)

2. Remove broken common/params/chunk_bundle.py

Imported TextChunkParamsByType (a symbol that does not exist), so it raised ImportError if ever loaded; nothing referenced it. Leftover of an incomplete refactor and a redundant third way to express chunking. (Report suggestion #24)

3. Unify the default LanceDB table name behind one constant

The default table name was spelled three different ways: nv-ingest (params, graph retriever, vdb operators/stage/bulk, harness config), nemo-retriever (ingest plan, query options, CLIs, harness resolution), and nemo_retriever (service vectordb + Helm). A default ingest and a default query could therefore silently target different tables — a latent bug.

Introduced common.vdb.DEFAULT_LANCEDB_TABLE_NAME = "nemo-retriever" (matching the user-facing ingest/query CLI defaults, which are unchanged) and routed every default through it: LanceDbParams, IngestStorageOptions, QueryStorageOptions, the graph Retriever fallback, the LanceDB operator/stage/bulk writers, the ingest & query CLIs, recall tooling, service VectorDbConfig/vectordb_app/Helm values, harness config/resolution, and benchmark_registry.DEFAULT_TABLE_NAME. (Report suggestion #6)

⚠️ Migration note: this changes the service vectordb default table name from nemo_retriever to nemo-retriever. Existing Helm/service deployments that relied on the old default should set serviceConfig.vectordb.tableName explicitly to keep reading pre-existing data. Evaluation tools that deliberately default to nv-ingest (pre-ingested benchmark tables) are intentionally left unchanged.

Intentionally NOT changed

• local_ingest_embed_backend default (vllm for ingest vs hf for query) — investigated and left as-is. This divergence is intentional (vLLM for batch-ingest throughput vs HF for low-latency single-query embedding); forcing them equal would likely regress one path.

Validation

• Base package installs; all changed modules import with no cycles; all defaults resolve to nemo-retriever.
• Passing: test_root_cli_workflow, test_root_query_cli, test_query_workflow_options, test_ingest_empty_validation, test_type_aliases, test_beir_evaluation, test_evaluation_retrievers, test_nv_ingest_vdb_operator, test_service_vectordb_app, test_service_vectordb_evidence_integration, test_helm_vectordb_embed_required, test_harness_agentic_eval, test_split_config* (170+ tests). GPU/[local]-extra tests were compile-checked only (torch not available in CI env).

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

  

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.