Adding lexical/semantic eval tools to Ollama Docker image

[tefirman]

Type of Change

• Version update (extends an existing image's bundled tooling)
• Documentation update

Description

Adds the Python dependencies needed by the WDL functional-eval pipeline (lexical/semantic similarity scoring and HuggingFace embedding loading) directly into the getwilds/ollama image so callers no longer have to pip install them at runtime. Previously, wilds-wdl-writer's wdl_writer/benchmarking.py (run via benchmarking.wdl) crashed on import because the image only shipped ollama and chromadb.

New packages, with pins matching a fresh install today:

• rapidfuzz==3.14.5 — fast string similarity for lexical eval
• sentence-transformers==5.5.1 — embedding models for semantic similarity
• llama-index-embeddings-huggingface==0.7.0 — the adapter benchmarking.py imports directly
• torch==2.12.0+cpu — installed first from https://download.pytorch.org/whl/cpu so the CUDA wheels (multi-GB) are skipped. Ollama owns the GPU; the embedding model runs on CPU. The smoke test asserts not torch.cuda.is_available() to catch regressions if a future change accidentally pulls in the CUDA build.

Applied identically to Dockerfile_0.21.0 and Dockerfile_latest. README updated: intro mentions lexical/semantic eval support, Image Details and Dockerfile Structure list the new packages, and citations were added for sentence-transformers and PyTorch.

Testing

How did you test these changes?

• make lint IMAGE=ollama — hadolint passes on both Dockerfiles.
• Built the image locally and confirmed it comes up cleanly. The expanded smoke test exercises every new import (rapidfuzz, torch + CPU-only assertion, sentence_transformers, from llama_index.embeddings.huggingface import HuggingFaceEmbedding) as part of the build.

Did the tests pass?

Yes — lint clean and local build succeeded.

Checklist

• Dockerfile follows naming convention (Dockerfile_X.Y.Z or Dockerfile_latest)
• All required OCI metadata labels are present and accurate
• README.md is included/updated in the tool directory
• Tested locally with make validate IMAGE=toolname (or manually built and verified)
• Image builds successfully for target platform(s)