New Architecture

[saksham-nexla]

---

[saksham-nexla]

Code review

Found 1 issue:

1. Hardcoded "gpt-4o" model name breaks non-OpenAI providers (bug in extract_simple and CLI commands)

The extract_simple() function defaults to model="gpt-4o" regardless of which provider is specified. This will cause runtime errors when users call extract_simple(document, schema, provider="anthropic") or any non-OpenAI provider without explicitly providing a model parameter. The model name "gpt-4o" is OpenAI-specific and will fail when passed to other providers like Anthropic, Google, Cohere, etc. The same bug appears in CLI commands suggest_schema.py and batch.py.

nextract/nextract/__init__.py

Lines 29 to 43 in 9a585af

	def extract_simple(
	document: str,
	schema: dict,
	provider: str,
	model: str | None = None,
	prompt: str | None = None,
	) -> Any:
	"""Simplest extraction helper using a default text pipeline."""
	provider_config = ProviderConfig(name=provider, model=model or "gpt-4o")
	extractor_config = ExtractorConfig(name="text", provider=provider_config)
	chunker_config = ChunkerConfig(name="semantic")
	plan = ExtractionPlan(extractor=extractor_config, chunker=chunker_config)
	pipeline = ExtractionPipeline(plan)
	return pipeline.extract(document=document, schema=schema, prompt=prompt)

nextract/nextract/cli/commands/suggest_schema.py

Lines 16 to 26 in 9a585af

	@app.command("suggest-schema")
	def cli_suggest_schema(
	samples: list[Path] = typer.Argument(..., exists=True, readable=True),
	prompt: str = typer.Option(..., "--prompt", "-p"),
	provider: str = typer.Option("openai", "--provider"),
	model: Optional[str] = typer.Option(None, "--model"),
	examples: Optional[Path] = typer.Option(None, "--examples"),
	output: Optional[Path] = typer.Option(None, "--output", "-o"),
	) -> None:
	provider_config = ProviderConfig(name=provider, model=model or "gpt-4o")
	generator = SchemaGenerator(provider=provider_config)

nextract/nextract/cli/commands/batch.py

Lines 20 to 38 in 9a585af

	@app.command("batch")
	def cli_batch(
	documents: list[Path] = typer.Argument(..., exists=True, readable=True),
	schema: Path = typer.Option(..., "--schema", "-s"),
	prompt: Optional[str] = typer.Option(None, "--prompt", "-p"),
	extractor: str = typer.Option("text", "--extractor"),
	provider: str = typer.Option("openai", "--provider"),
	model: Optional[str] = typer.Option(None, "--model"),
	chunker: str = typer.Option("semantic", "--chunker"),
	max_workers: int = typer.Option(4, "--max-workers"),
	include_extra: bool = typer.Option(False, "--include-extra"),
	num_passes: int = typer.Option(1, "--num-passes"),
	enable_suggestions: bool = typer.Option(False, "--enable-suggestions"),
	) -> None:
	schema_obj = _load_schema(schema)
	provider_config = ProviderConfig(name=provider, model=model or "gpt-4o")
	extractor_config = ExtractorConfig(name=extractor, provider=provider_config)
	chunker_config = ChunkerConfig(name=chunker)

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[devin-ai-integration]

Devin Review found 3 potential issues.

View 7 additional findings in Devin Review.

[devin-ai-integration]

🔴 Excel (.xls/.xlsx) reclassification makes text extraction dead code, routes files to PDF conversion instead

Moving .xls and .xlsx from _TEXTUAL_EXTS to _UNSUPPORTED_BUT_ACCEPTABLE_AS_BINARY in nextract/mimetypes_map.py causes is_textual() to return False for these extensions. However, _prepare_single_file in nextract/files.py:308-330 still has specialized Excel→TSV/CSV text extraction logic gated by is_textual(path). Since that guard now evaluates to False, the entire Excel text-extraction branch (.xlsx→TSV via _xlsx_to_text, .xls→CSV via _xls_to_text_via_cli) is unreachable dead code. Instead, Excel files now fall through to is_office_binary() at nextract/files.py:383, which attempts Office→PDF conversion — a much worse outcome for spreadsheets that were previously extracted as structured text.

Prompt for agents

The root issue is in nextract/mimetypes_map.py lines 36-41: .xls and .xlsx were moved from _TEXTUAL_EXTS to _UNSUPPORTED_BUT_ACCEPTABLE_AS_BINARY. But nextract/files.py _prepare_single_file (lines 308-330) has specialized Excel text-extraction code gated by is_textual(path), which now always returns False for these extensions. Either: (1) Move .xls/.xlsx back to _TEXTUAL_EXTS and remove them from _UNSUPPORTED_BUT_ACCEPTABLE_AS_BINARY, OR (2) Update _prepare_single_file in files.py to check for Excel extensions before the is_textual() guard (e.g. add an explicit check like `if path.suffix.lower() in {".xlsx", ".xls"}:` before line 308).

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Multipass 'union' merge strategy converts scalar fields to lists, breaking schema validation

The new union merge strategy in _merge_results wraps all field values into a list via merged_list.append(v). For scalar fields (strings, numbers), this converts e.g. {"name": "John"} across two passes into {"name": ["John", "John"]} instead of keeping it as {"name": "John"}. This breaks JSON Schema validation for any non-array field. The old behavior treated union as equivalent to first_non_empty (take first non-empty value), which preserved scalar types.

Example of the breakage

Given two passes both returning {"name": "John", "total": 500} with a schema {"properties": {"name": {"type": "string"}, "total": {"type": "number"}}}, the union merge now produces {"name": ["John", "John"], "total": [500, 500]} — both fields are now lists, violating the schema.

Suggested change

	if strategy == "union":
	if values:
	# For list values, merge all lists
	merged_list = []
	for v in values:
	if isinstance(v, list):
	merged_list.extend(v)
	elif v:
	merged_list.append(v)
	merged[field_name] = merged_list if merged_list else values[0]
	if strategy == "union":
	if values:
	# For list values, merge all lists; for scalars, take first
	if any(isinstance(v, list) for v in values):
	merged_list = []
	for v in values:
	if isinstance(v, list):
	merged_list.extend(v)
	elif v is not None:
	merged_list.append(v)
	merged[field_name] = merged_list if merged_list else values[0]
	else:
	merged[field_name] = values[0]

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Multipass 'union' merge drops falsy values like 0 and False

In the union merge strategy at nextract/multipass.py:288, the condition elif v: uses a truthiness check. This silently drops legitimate falsy values such as 0, 0.0, and False from the merged output. For example, a field with "total": 0 would pass the earlier filter at line 277 (value is not None and value != "" and value != []) but then be skipped by the elif v: guard, causing data loss.

Suggested change

	elif v:
	merged_list.append(v)
	elif v is not None:
	merged_list.append(v)

---

Was this helpful? React with 👍 or 👎 to provide feedback.