Persist and reload FAISSRetriever and InMemoryRetriever indexes

[adaumsilva]

Description

Only ChromaRetriever survives a process restart. FAISSRetriever and InMemoryRetriever must re-embed the whole corpus every run. Add:

retriever.save("index_dir/")
retriever = FAISSRetriever.load("index_dir/")

Motivation

Embedding is the slow, expensive step (API cost or GPU time). Losing the index on exit makes the two fastest retrievers impractical beyond demos.

Acceptance criteria

• save(path: str | Path) -> None and @classmethod load(path) -> Self on both classes
• FAISSRetriever.save writes index.faiss (faiss.write_index) + chunks.json (ids, content, metadata) + meta.json (dimension, m, ef_construction, ef_search, format version)
• InMemoryRetriever.save writes matrix.npy + chunks.json
• Use JSON, not pickle, for chunk data (safe to load untrusted files); np.load(..., allow_pickle=False)
• load() on a missing/corrupt directory raises RetrieverError; format-version mismatch raises RetrieverError with guidance
• Round-trip tests with tmp_path: save → load → identical len() and identical retrieve() results
• Docstrings + examples/README.md snippet; CHANGELOG.md updated

Files to touch

• ragframework/retriever/faiss.py
• ragframework/retriever/in_memory.py
• tests/test_retriever/test_faiss.py, tests/test_retriever/test_in_memory.py
• examples/README.md

Resources

• faiss.write_index / read_index
• numpy.save / numpy.load

Estimated effort: Medium (half day)

[Guru2907]

Hi @adaumsilva, I'd like to work on this issue. I'll implement persistence and reload support for FAISSRetriever and InMemoryRetriever, along with the requested round-trip tests and documentation updates.

[adaumsilva]

Assigned to you @Guru2907