fix: Drop the Dead NLTK Download That Breaks Image Builds

[danny-avila]

Summary

Every image build on main has failed since 52989197 bumped unstructured from 0.18.32 to 0.27.8. Both Dockerfiles run a step that downloads NLTK data, and it now dies with ModuleNotFoundError: No module named 'nltk' (failed job). Nothing is being published to librechat-rag-api-dev or -dev-lite until this lands.

unstructured 0.18.32 declared nltk as a dependency; 0.27.8 does not. That is not an oversight on their side: 0.27.8 no longer uses NLTK at all. It ships its own sent_tokenize, word_tokenize and pos_tag in unstructured/nlp/tokenize.py, and none of its 138 modules mention NLTK. This repository's own code never imported it either.

So the step is dead weight rather than a missing dependency. This removes it, along with the NLTK_DATA variable, from both Dockerfile and Dockerfile.lite, instead of reinstalling NLTK and downloading data nothing reads.

How it works

 RUN pip install --no-cache-dir -r requirements.lite.txt
-ENV NLTK_DATA=/app/nltk_data
-
-# Download standard NLTK data, to prevent unstructured from downloading packages at runtime
-RUN python -c "import nltk, sys; packages = ('punkt_tab', ...)"
-
 # Disable Unstructured analytics

Same change in both files. The images also get smaller, since they no longer carry the punkt and tagger data.

Testing

• Checked PyPI metadata: unstructured 0.18.32 requires_dist includes nltk; 0.27.8 lists none.
• Extracted the 0.27.8 wheel and searched every module: no import nltk, no reference to it at all.
• Searched app/ here: no NLTK usage.
• The image build is the real check. It runs on merge to main, not on pull requests, so the first build after merge is what confirms it.