fix(synthesizer): avoid UnboundLocalError when a document yields 0 chunks - #2849

Summary

ContextGenerator.validate_chunk_size is meant to raise an informative ValueError when a source document produces fewer chunks than the requested number of contexts. Instead, when a document yields 0 chunks (e.g. an empty / whitespace-only / unparseable file), it raises a confusing UnboundLocalError, hiding the real cause from the user.

Reproduction

from deepeval.synthesizer.chunking.context_generator import ContextGenerator

class FakeCollection:          # stand-in Chroma collection
    def count(self): return 0  # document produced 0 chunks

class StubEmbedder:
    def embed_text(self, x): return [0.0]
    async def a_embed_text(self, x): return [0.0]
    def embed_texts(self, xs): return [[0.0] for _ in xs]
    async def a_embed_texts(self, xs): return [[0.0] for _ in xs]

gen = ContextGenerator(document_paths=["doc.txt"], embedder=StubEmbedder(), model=<any model>)
gen.validate_chunk_size(min_contexts_per_source_file=1, collection=FakeCollection())

On main this raises:

UnboundLocalError: cannot access local variable 'adjust_chunk_size' where it is not associated with a value

instead of the intended ValueError describing the 0-chunk problem.

Root cause

In validate_chunk_size, adjust_chunk_size and adjust_overlap are only assigned inside the num_chunks > 0 and min_contexts_per_source_file > 1 and self.chunk_overlap > 0 branches, but the final combined-suggestion check evaluates them unconditionally:

if adjust_chunk_size or adjust_overlap:   # both unbound when num_chunks == 0

When a document yields 0 chunks, none of those branches run, so the names are never bound and Python raises UnboundLocalError.

Fix

• Initialize adjust_chunk_size = False and adjust_overlap = False before the suggestion branches, so the combined check is always safe.
• Only emit the "You have the following options:" header when there is at least one concrete suggestion, so the 0-chunk message no longer dangles with an empty options list.

After the fix, a 0-chunk document produces a clean, actionable error:

Impossible to generate 1 contexts from a document with 0 chunks.

and the existing multi-suggestion messages are unchanged for the normal shortfall cases.

Tests

Adds four regression tests in tests/test_core/test_synthesizer/test_context_generator.py (no API key required — a stub LLM is injected so ContextGenerator constructs offline):

• test_validate_chunk_size_zero_chunks_raises_value_error — 0 chunks, min_contexts=1
• test_validate_chunk_size_zero_chunks_with_overlap_raises_value_error — 0 chunks, min_contexts>1, non-zero overlap
• test_validate_chunk_size_insufficient_chunks_still_lists_options — normal shortfall path keeps the full suggestion list
• test_validate_chunk_size_passes_when_enough_chunks — no error when there are enough chunks

The two zero-chunk tests fail on main with UnboundLocalError and pass with this change.

Notes

• Scope is intentionally minimal and does not touch the min_contexts_per_document wording that fix(synthesizer): correct parameter name in validate_chunk_size error message #2759 corrects (that is a separate line), so there is no conflict.
• Formatted with black.