Llm

pipelines.extraction.pdf.llm

Summary

LLM splitter variant of the pdf family (impl-layer, bank-agnostic).

---

LLMSplitterPort is a faithful port of TransactionSplitterLLMAdaptor (domain/extraction/adaptors/txn_blocks.py) using SyncOpenRouterPool. It swaps into the pdf chain exactly where ImplSplitterPort sits, so the rest of the tree (raw_lines → txn_blocks → txn_dicts → raw_expense) is unchanged.

LLMPdfExtractionPipeline re-pins only the txn_blocks.splitter slot.

Classes

LLMPdfExtractionPipeline

Bases: ImplPDFExtractionPipeline

PDF root with the LLM splitter swapped into the txn-blocks stage.

Attributes:

Name	Type	Description
txn_blocks	LLMTxnBlocksPort	Txn-blocks stage with LLM splitter.

Functions

run

run(source)

Run the pdf extraction chain end-to-end.

Parameters:

Name	Type	Description	Default
source		Raw extraction source.	required

Returns:

Type	Description
	List of raw expense records.

LLMSplitterPort

LLMSplitterPort(pool=None, **kwargs)

Bases: SplitterPort

LLM-backed splitter that groups lines into transactions via prompts.

Attributes:

Name	Type	Description
DATE_RE	Pattern[str]	Compiled regex matching leading DD-MM-YYYY dates.
number	ImplNumberPort	Number extraction leaf.
date	ImplDatePort	Date parsing leaf.
GROUP_PROMPT	str	Prompt grouping indices into transactions.
RESOLVE_PROMPT	str	Prompt resolving cross-batch conflicts.
MAX_CONFLICT_LOOPS	int	Max conflict resolution loops.
BATCH_MERGE	int	Date lines per batch.

Initialize the splitter with an LLM pool.

Parameters:

Name	Type	Description	Default
pool		LLM pool; defaults to a new SyncOpenRouterPool.	None
**kwargs		Passed through to the base class.	{}

Functions

clean_raw_lines staticmethod

clean_raw_lines(raw_lines: List[str]) -> List[str]

Clean raw lines: strip trailing noise and drop Page lines.

Parameters:

Name	Type	Description	Default
raw_lines	List[str]	Raw OCR lines.	required

Returns:

Type	Description
List[str]	List[str]: Cleaned lines.

split_blocks

split_blocks(
    lines,
    *,
    has_header=False,
    min_numbers=3,
    detect_date_ranges=False
)

Group lines into transaction blocks via the LLM.

Parameters:

Name	Type	Description	Default
lines		Raw OCR lines.	required
has_header		Whether a header precedes the rows.	False
min_numbers		Minimum numbers for a block.	3
detect_date_ranges		Detect date-range lines.	False

Returns:

Type	Description
	List of transaction block texts.

strip_headers staticmethod

strip_headers(raw_lines: List[str]) -> List[str]

Drop header lines preceding the first date line.

Parameters:

Name	Type	Description	Default
raw_lines	List[str]	Raw OCR lines.	required

Returns:

Type	Description
List[str]	List[str]: Lines with headers removed.

LLMTxnBlocksPort

Bases: ImplTxnBlocksPort

Txn-blocks stage re-pinning the splitter to the LLM variant.

Attributes:

Name	Type	Description
splitter	LLMSplitterPort	LLM-backed splitter.

Functions

extract

extract(raw_lines)

Extract txn blocks keyed by opening balance.

Parameters:

Name	Type	Description	Default
raw_lines		Raw OCR text lines.	required

Returns:

Type	Description
	Dict[float, List[str]]: Opening balance → block texts.