[Harithalakshmikt]

Adds a local RAG example using LangChain, FAISS and google/flan-t5-base,
with check_with_evidence() and an NLI-based compare_fn
(cross-encoder/nli-deberta-v3-base).

Tested model: google/flan-t5-base (local, no API key).
Includes a deliberately wrong answer (BLOCKED) and a correct control.
RESULTS.md reports the failures honestly: the compare_fn is over-strict
and blocks some correct answers.

Related to #2

[Harithalakshmikt]

Hi @Vedantgitbot, I've opened a PR for this example: (#22)
)
It uses LangChain + FAISS + google/flan-t5-base (local, no API key) with check_with_evidence() and an NLI-based compare_fn. It includes a deliberately wrong answer (BLOCKED) and a correct control answer. RESULTS.md documents the failures honestly, including false blocks from the over-strict compare_fn.

Happy to make changes if you'd like any.

[Vedantgitbot]

Thanks @Harithalakshmikt Great writeup, especially the NLI comparator and false-block findings. Before merging, please fix the docstring mismatch and use pip install boothpy instead of -e .. Optional: try chunk_size=100 and add a default question for reproducibility. #18 on compare_fn examples could be a great next step!

[Harithalakshmikt]

Thanks @Vedantgitbot! Pushed the changes: the docstring now matches the script (two hardcoded checks, one wrong and one correct), the install line is pip install boothpy, and I added a default question (press Enter). I left chunk_size unchanged because changing it would invalidate the recorded results; happy to try it as a follow-up.