data-science-document-ai 1.43.6__tar.gz → 1.45.2__tar.gz

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: data-science-document-ai
-Version: 1.43.6
+Version: 1.45.2
 Summary: "Document AI repo for data science"
 Author: Naomi Nguyen
 Author-email: naomi.nguyen@forto.com

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "data-science-document-ai"
-version = "1.43.6"
+version = "1.45.2"
 description = "\"Document AI repo for data science\""
 authors = ["Naomi Nguyen <naomi.nguyen@forto.com>", "Kumar Rajendrababu <kumar.rajendrababu@forto.com>", "Igor Tonko <igor.tonko@forto.com>", "Osman Demirel <osman.demirel@forto.com>"]
 packages = [

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/constants.py

@@ -23,9 +23,12 @@ project_parameters = {
     "invoice_classification_lookup": "invoice_classification.json",
     "reverse_charge_sentence_lookup": "reverse_charge_sentences.json",
     # Fuzzy logic params
-    "fuzzy_threshold_item_code": 70,
+    "fuzzy_threshold_item_code": 90,
     "fuzzy_threshold_reverse_charge": 80,
     "fuzzy_threshold_invoice_classification": 70,
+    # Chunking params
+    "chunk_size": 1,  # page (do not change this without changing the page number logic)
+    "chunk_after": 10,  # pages
     # Big Query
     "g_ai_gbq_db_schema": "document_ai",
     "g_ai_gbq_db_table_out": "document_ai_api_calls_v1",

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/excel_processing.py

@@ -4,8 +4,6 @@ import logging
 
 from ddtrace import tracer
 
-from src.postprocessing.common import llm_prediction_to_tuples
-
 logger = logging.getLogger(__name__)
 
 import asyncio
@@ -78,6 +76,7 @@ async def extract_data_from_excel(
             "bundeskasse",
             "commercialInvoice",
             "packingList",
+            "bookingConfirmation",
         ]
         else generate_schema_structure(params, input_doc_type)
     )

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/io.py

@@ -156,4 +156,27 @@ def download_dir_from_bucket(bucket, directory_cloud, directory_local) -> bool:
     return result
 
 
+def bq_logs(data_to_insert, params):
+    """Insert logs into Google BigQuery.
+
+    Args:
+        data_to_insert (list): The data to insert into BigQuery.
+        params (dict): The parameters dictionary.
+    """
+    # Use the pre-initialized BigQuery client
+    bq_client = params["bq_client"]
+    # Get the table string
+    table_string = f"{params['g_ai_project_name']}.{params['g_ai_gbq_db_schema']}.{params['g_ai_gbq_db_table_out']}"
+
+    logger.info(f"Log table: {table_string}")
+    # Insert the rows into the table
+    insert_logs = bq_client.insert_rows_json(table_string, data_to_insert)
+
+    # Check if there were any errors inserting the rows
+    if not insert_logs:
+        logger.info("New rows have been added.")
+    else:
+        logger.info("Errors occurred while inserting rows: ", insert_logs)
+
+
 # type: ignore

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/pdf_processing.py

@@ -36,6 +36,7 @@ from src.utils import (
     get_pdf_page_count,
     get_processor_name,
     run_background_tasks,
+    split_pdf_into_chunks,
     transform_schema_strings,
     validate_based_on_schema,
 )
@@ -195,15 +196,11 @@ async def process_file_w_llm(params, file_content, input_doc_type, llm_client):
         result (dict): The structured data extracted from the document, formatted as JSON.
     """
     # Bundeskasse invoices contains all the required information in the first 3 pages.
-    file_content = (
-        extract_top_pages(file_content, num_pages=5)
-        if input_doc_type == "bundeskasse"
-        else file_content
-    )
-    number_of_pages = get_pdf_page_count(file_content)
+    if input_doc_type == "bundeskasse":
+        file_content = extract_top_pages(file_content, num_pages=5)
 
-    # convert file_content to required document
-    document = llm_client.prepare_document_for_gemini(file_content)
+    number_of_pages = get_pdf_page_count(file_content)
+    logger.info(f"processing {input_doc_type} with {number_of_pages} pages...")
 
     # get the schema placeholder from the Doc AI and generate the response structure
     response_schema = (
@@ -215,26 +212,28 @@ async def process_file_w_llm(params, file_content, input_doc_type, llm_client):
             "bundeskasse",
             "commercialInvoice",
             "packingList",
+            "bookingConfirmation",
         ]
         else generate_schema_structure(params, input_doc_type)
     )
 
     carrier = "other"
-    if (
-        "preprocessing" in prompt_library.library.keys()
-        and "carrier" in prompt_library.library["preprocessing"].keys()
-        and input_doc_type
-        in prompt_library.library["preprocessing"]["carrier"]["placeholders"].keys()
-    ):
-        carrier_schema = prompt_library.library["preprocessing"]["carrier"][
-            "placeholders"
-        ][input_doc_type]
+    carrier_schema = (
+        prompt_library.library.get("preprocessing", {})
+        .get("carrier", {})
+        .get("placeholders", {})
+        .get(input_doc_type)
+    )
 
+    if carrier_schema:
         carrier_prompt = prompt_library.library["preprocessing"]["carrier"]["prompt"]
         carrier_prompt = carrier_prompt.replace(
             "DOCUMENT_TYPE_PLACEHOLDER", input_doc_type
         )
 
+        # convert file_content to required document
+        document = llm_client.prepare_document_for_gemini(file_content)
+
         # identify carrier for customized prompting
         carrier = await identify_carrier(
             document,
@@ -244,37 +243,115 @@ async def process_file_w_llm(params, file_content, input_doc_type, llm_client):
             doc_type=input_doc_type,
         )
 
-    if input_doc_type == "bookingConfirmation":
-        response_schema = prompt_library.library[input_doc_type][carrier][
-            "placeholders"
-        ]
-
+    # Select prompt
     if (
-        input_doc_type in prompt_library.library.keys()
-        and carrier in prompt_library.library[input_doc_type].keys()
+        input_doc_type not in prompt_library.library
+        or carrier not in prompt_library.library[input_doc_type]
     ):
-        # get the related prompt from predefined prompt library
-        prompt = prompt_library.library[input_doc_type][carrier]["prompt"]
+        return {}
 
-        # Update schema to extract value-page_number pairs
-        if number_of_pages > 1:
-            response_schema = transform_schema_strings(response_schema)
+    # get the related prompt from predefined prompt library
+    prompt = prompt_library.library[input_doc_type][carrier]["prompt"]
 
-            # Update the prompt to instruct LLM to include page numbers
-            prompt += "\nFor each field, provide the page number where the information was found. The page numbering starts from 0."
+    # Add page-number extraction for moderately large docs
+    use_chunking = number_of_pages >= params["chunk_after"]
 
-        # generate the result with LLM (gemini)
-        result = await llm_client.get_unified_json_genai(
-            prompt=prompt,
-            document=document,
-            response_schema=response_schema,
-            doc_type=input_doc_type,
+    # Update schema and prompt to extract value-page_number pairs
+    if not use_chunking and number_of_pages > 1:
+        response_schema = transform_schema_strings(response_schema)
+        prompt += "\nFor each field, provide the page number where the information was found. The page numbering starts from 0."
+
+    tasks = []
+    # Process in chunks if number of pages exceeds threshold and Process all chunks concurrently
+    for chunk in (
+        split_pdf_into_chunks(file_content, chunk_size=params["chunk_size"])
+        if use_chunking
+        else [file_content]
+    ):
+        tasks.append(
+            process_chunk_with_retry(
+                chunk, prompt, response_schema, llm_client, input_doc_type
+            )
         )
 
-        result = llm_prediction_to_tuples(result, number_of_pages)
+    results = await asyncio.gather(*tasks, return_exceptions=True)
 
-        return result
-    return {}
+    if use_chunking:
+        return merge_llm_results(results, response_schema)
+    else:
+        return llm_prediction_to_tuples(results[0], number_of_pages=number_of_pages)
+
+
+async def process_chunk_with_retry(
+    chunk_content, prompt, response_schema, llm_client, input_doc_type, retries=2
+):
+    """Process a chunk with retries in case of failure."""
+    for attempt in range(1, retries + 1):
+        try:
+            return await process_chunk(
+                chunk_content=chunk_content,
+                prompt=prompt,
+                response_schema=response_schema,
+                llm_client=llm_client,
+                input_doc_type=input_doc_type,
+            )
+        except Exception as e:
+            logger.error(f"Chunk failed on attempt {attempt}: {e}")
+            if attempt == retries:
+                raise
+            await asyncio.sleep(1)  # small backoff
+
+
+async def process_chunk(
+    chunk_content, prompt, response_schema, llm_client, input_doc_type
+):
+    """Process a chunk with Gemini."""
+    document = llm_client.prepare_document_for_gemini(chunk_content)
+    return await llm_client.get_unified_json_genai(
+        prompt=prompt,
+        document=document,
+        response_schema=response_schema,
+        doc_type=input_doc_type,
+    )
+
+
+def merge_llm_results(results, response_schema):
+    """Merge LLM results from multiple chunks."""
+    merged = {}
+    for i, result in enumerate(results):
+        if not isinstance(result, dict):
+            continue
+        # Add page number to all values coming from this chunk
+        result = llm_prediction_to_tuples(result, number_of_pages=1, page_number=i)
+
+        # Merge the result into the final merged dictionary
+        for key, value in result.items():
+            field_type = (
+                response_schema["properties"].get(key, {}).get("type", "").upper()
+            )
+
+            if key not in merged:
+                if field_type == "ARRAY":
+                    # append the values as a list
+                    merged[key] = (
+                        value if isinstance(value, list) else ([value] if value else [])
+                    )
+                else:
+                    merged[key] = value
+                continue
+
+            if field_type == "ARRAY":
+                # append list contents across chunks
+                if isinstance(value, list):
+                    merged[key].extend(value)
+                else:
+                    merged[key].append(value)
+
+            # take first non-null value only
+            if merged[key] in (None, "", [], {}):
+                merged[key] = value
+
+    return merged
 
 
 async def extract_data_from_pdf_w_llm(params, input_doc_type, file_content, llm_client):

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/postprocessing/common.py

@@ -12,7 +12,7 @@ from src.constants import formatting_rules
 from src.io import logger
 from src.postprocessing.postprocess_partner_invoice import process_partner_invoice
 from src.prompts.prompt_library import prompt_library
-from src.utils import get_tms_mappings
+from src.utils import batch_fetch_all_mappings, get_tms_mappings
 
 tms_domain = os.environ["TMS_DOMAIN"]
 
@@ -134,8 +134,11 @@ def extract_number(data_field_value):
         formatted_value: string
 
     """
+    # Remove container size pattern like 20FT, 40HC, etc from 1 x 40HC
+    value = remove_unwanted_patterns(data_field_value)
+
     formatted_value = ""
-    for c in data_field_value:
+    for c in value:
         if c.isnumeric() or c in [",", ".", "-"]:
             formatted_value += c
 
@@ -320,9 +323,12 @@ def remove_unwanted_patterns(lineitem: str):
     lineitem = lineitem.replace("HIGH CUBE", "")
 
     # Remove container size e.g., 20FT, 40HC, etc.
-    lineitem = re.sub(
-        r"\b(20|22|40|45)(FT|HC|DC|HD|GP|OT|RF|FR|TK|DV)?\b", "", lineitem
-    ).strip()
+    pattern = [
+        f"{s}{t}"
+        for s in ("20|22|40|45".split("|"))
+        for t in ("FT|HC|DC|HD|GP|OT|RF|FR|TK|DV".split("|"))
+    ]
+    lineitem = re.sub(r"|".join(pattern), "", lineitem, flags=re.IGNORECASE).strip()
 
     return lineitem
 
@@ -372,18 +378,45 @@ def clean_item_description(lineitem: str, remove_numbers: bool = True):
     return re.sub(r"\s{2,}", " ", lineitem).strip()
 
 
-async def format_label(entity_k, entity_value, document_type_code, params, mime_type):
+async def format_label(
+    entity_k,
+    entity_value,
+    document_type_code,
+    params,
+    mime_type,
+    container_map,
+    terminal_map,
+    depot_map,
+):
     llm_client = params["LlmClient"]
     if isinstance(entity_value, dict):  # if it's a nested entity
         format_tasks = [
-            format_label(sub_k, sub_v, document_type_code, params, mime_type)
+            format_label(
+                sub_k,
+                sub_v,
+                document_type_code,
+                params,
+                mime_type,
+                container_map,
+                terminal_map,
+                depot_map,
+            )
             for sub_k, sub_v in entity_value.items()
         ]
         return entity_k, {k: v for k, v in await asyncio.gather(*format_tasks)}
     if isinstance(entity_value, list):
         format_tasks = await asyncio.gather(
             *[
-                format_label(entity_k, sub_v, document_type_code, params, mime_type)
+                format_label(
+                    entity_k,
+                    sub_v,
+                    document_type_code,
+                    params,
+                    mime_type,
+                    container_map,
+                    terminal_map,
+                    depot_map,
+                )
                 for sub_v in entity_value
             ]
         )
@@ -405,13 +438,13 @@ async def format_label(entity_k, entity_value, document_type_code, params, mime_
         )
 
     elif (entity_key == "containertype") or (entity_key == "containersize"):
-        formatted_value = get_tms_mappings(entity_value, "container_types")
+        formatted_value = container_map.get(entity_value)
 
     elif check_formatting_rule(entity_k, document_type_code, "terminal"):
-        formatted_value = get_tms_mappings(entity_value, "terminals")
+        formatted_value = terminal_map.get(entity_value)
 
     elif check_formatting_rule(entity_k, document_type_code, "depot"):
-        formatted_value = get_tms_mappings(entity_value, "depots")
+        formatted_value = depot_map.get(entity_value)
 
     elif entity_key.startswith(("eta", "etd", "duedate", "issuedate", "servicedate")):
         try:
@@ -507,7 +540,8 @@ async def get_port_code_ai(port: str, llm_client, doc_type=None):
     """Get port code using AI model."""
     port_llm = await get_port_code_llm(port, llm_client, doc_type=doc_type)
 
-    return get_tms_mappings(port, "ports", port_llm)
+    result = await get_tms_mappings(port, "ports", port_llm)
+    return result.get(port, None)
 
 
 async def get_port_code_llm(port: str, llm_client, doc_type=None):
@@ -598,6 +632,74 @@ def decimal_convertor(value, quantity=False):
     return value
 
 
+async def collect_mapping_requests(entity_value, document_type_code):
+    """Collect all unique container types, terminals, and depots from the entity value."""
+    # Sets to store unique values
+    container_types = set()
+    terminals = set()
+    depots = set()
+
+    def walk(key, value):
+        key_lower = key.lower()
+
+        # nested dict
+        if isinstance(value, dict):
+            for k, v in value.items():
+                walk(k, v)
+
+        # list of values
+        elif isinstance(value, list):
+            for item in value:
+                walk(key, item)
+
+        # leaf node
+        else:
+            if key_lower in ("containertype", "containersize"):
+                # Take only "20DV" from ('20DV', 0) if it's a tuple
+                container_types.add(value[0]) if isinstance(
+                    value, tuple
+                ) else container_types.add(value)
+
+            elif check_formatting_rule(key, document_type_code, "terminal"):
+                terminals.add(value[0]) if isinstance(value, tuple) else terminals.add(
+                    value
+                )
+
+            elif check_formatting_rule(key, document_type_code, "depot"):
+                depots.add(value[0]) if isinstance(value, tuple) else depots.add(value)
+
+    walk("root", entity_value)
+
+    return container_types, terminals, depots
+
+
+async def format_all_labels(entity_data, document_type_code, params, mime_type):
+    """Format all labels in the entity data using cached mappings."""
+    # Collect all mapping values needed
+    container_req, terminal_req, depot_req = await collect_mapping_requests(
+        entity_data, document_type_code
+    )
+
+    # Batch fetch mappings
+    container_map, terminal_map, depot_map = await batch_fetch_all_mappings(
+        container_req, terminal_req, depot_req
+    )
+
+    # Format labels using cached mappings
+    _, result = await format_label(
+        "root",
+        entity_data,
+        document_type_code,
+        params,
+        mime_type,
+        container_map,
+        terminal_map,
+        depot_map,
+    )
+
+    return _, result
+
+
 async def format_all_entities(result, document_type_code, params, mime_type):
     """Format the entity values in the result dictionary."""
     # Since we treat `customsInvoice` same as `partnerInvoice`
@@ -613,13 +715,13 @@ async def format_all_entities(result, document_type_code, params, mime_type):
         return {}
 
     # Format all entities recursively
-    _, aggregated_data = await format_label(
-        None, result, document_type_code, params, mime_type
+    _, aggregated_data = await format_all_labels(
+        result, document_type_code, params, mime_type
     )
 
     # Process partner invoice on lineitem mapping and reverse charge sentence
     if document_type_code in ["partnerInvoice", "bundeskasse"]:
-        process_partner_invoice(params, aggregated_data, document_type_code)
+        await process_partner_invoice(params, aggregated_data, document_type_code)
 
     logger.info("Data Extraction completed successfully")
     return aggregated_data
@@ -651,41 +753,46 @@ def remove_stop_words(lineitem: str):
     )
 
 
-def llm_prediction_to_tuples(llm_prediction, number_of_pages=-1):
+def llm_prediction_to_tuples(llm_prediction, number_of_pages=-1, page_number=None):
     """Convert LLM prediction dictionary to tuples of (value, page_number)."""
-
     # If only 1 page, simply pair each value with page number 0
     if number_of_pages == 1:
+        effective_page = 0 if page_number is None else page_number
         if isinstance(llm_prediction, dict):
             return {
-                k: llm_prediction_to_tuples(v, number_of_pages)
+                k: llm_prediction_to_tuples(
+                    v, number_of_pages, page_number=effective_page
+                )
                 for k, v in llm_prediction.items()
             }
         elif isinstance(llm_prediction, list):
             return [
-                llm_prediction_to_tuples(v, number_of_pages) for v in llm_prediction
+                llm_prediction_to_tuples(v, number_of_pages, page_number=effective_page)
+                for v in llm_prediction
             ]
         else:
-            return (llm_prediction, 0) if llm_prediction else None
+            return (llm_prediction, effective_page) if llm_prediction else None
 
     # logic for multi-page predictions
     if isinstance(llm_prediction, dict):
         if "page_number" in llm_prediction.keys() and "value" in llm_prediction.keys():
             if llm_prediction["value"]:
                 try:
-                    page_number = int(llm_prediction["page_number"])
+                    _page_number = int(llm_prediction["page_number"])
                 except:  # noqa: E722
-                    page_number = -1
-                return (llm_prediction["value"], page_number)
+                    _page_number = -1
+                return (llm_prediction["value"], _page_number)
             return None
 
         for key, value in llm_prediction.items():
             llm_prediction[key] = llm_prediction_to_tuples(
-                llm_prediction.get(key, value), number_of_pages
+                llm_prediction.get(key, value), number_of_pages, page_number
             )
 
     elif isinstance(llm_prediction, list):
         for i, item in enumerate(llm_prediction):
-            llm_prediction[i] = llm_prediction_to_tuples(item, number_of_pages)
+            llm_prediction[i] = llm_prediction_to_tuples(
+                item, number_of_pages, page_number
+            )
 
     return llm_prediction

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/postprocessing/postprocess_partner_invoice.py

@@ -1,7 +1,5 @@
 """This module contains the postprocessing functions for the partner invoice."""
-from concurrent.futures import ThreadPoolExecutor
-
-from fuzzywuzzy import fuzz
+from rapidfuzz import fuzz, process
 
 from src.io import logger
 from src.utils import get_tms_mappings
@@ -136,7 +134,7 @@ def update_recipient_and_vendor(aggregated_data, is_recipient_forto):
     ] = "Dasbachstraße 15, 54292 Trier, Germany"
 
 
-def process_partner_invoice(params, aggregated_data, document_type_code):
+async def process_partner_invoice(params, aggregated_data, document_type_code):
     """Process the partner invoice data."""
     # Post process bundeskasse invoices
     if document_type_code == "bundeskasse":
@@ -160,27 +158,80 @@ def process_partner_invoice(params, aggregated_data, document_type_code):
         reverse_charge_info["formattedValue"] = reverse_charge_value
         reverse_charge = aggregated_data.pop("reverseChargeSentence", None)
 
-    # Process each line item
-    for line_item in line_items:
-        if line_item.get("lineItemDescription", None) is not None:
-            line_item["itemCode"] = associate_forto_item_code(
-                line_item["lineItemDescription"]["formattedValue"],
-                params,
-            )
+    # Process everything in one go
+    processed_items = await process_line_items_batch(params, line_items, reverse_charge)
 
-            # Add page number for the consistency
-            line_item["itemCode"]["page"] = line_item["lineItemDescription"]["page"]
+    # Update your main data structure
+    aggregated_data["lineItem"] = processed_items
 
-        if reverse_charge:
-            # Distribute reverseChargeSentence to all line items
-            line_item["reverseChargeSentence"] = reverse_charge
-            line_item["reverseChargeSentence"]["page"] = reverse_charge["page"]
 
+async def process_line_items_batch(
+    params: dict, line_items: list[dict], reverse_charge=None
+):
+    """
+    Processes all line items efficiently using a "Split-Apply-Combine" strategy.
+    """
+    # To store items that need external API lookup
+    pending_line_items = {}
+
+    # Check Fuzzy Matching
+    logger.info(f"Mapping line item codes with Fuzzy matching....")
+    for i, item in enumerate(line_items):
+        description_obj = item.get("lineItemDescription")
+
+        if not description_obj or not description_obj.get("formattedValue"):
+            continue
+        # Get the formatted description text
+        desc = description_obj["formattedValue"]
+
+        # Find Fuzzy Match
+        matched_code = find_matching_lineitem(
+            desc,
+            params["lookup_data"]["item_code"],
+            params["fuzzy_threshold_item_code"],
+        )
+
+        if matched_code:
+            # Set the code to the line item
+            item["itemCode"] = {
+                "documentValue": desc,
+                "formattedValue": matched_code,
+                "page": description_obj.get("page"),
+            }
+        else:
+            # Store for batch API call
+            pending_line_items[i] = desc
+
+    # Batch API Call for Embedding lookups
+    if pending_line_items:
+        values_to_fetch = list(set(pending_line_items.values()))
+        logger.info(f"Mapping {len(values_to_fetch)} line items from Embedding API...")
+
+        # Await the batch response {"desc1": "code1", "desc2": "code2"}
+        api_results = await get_tms_mappings(
+            input_list=values_to_fetch, embedding_type="line_items"
+        )
+
+        # Merge API results back into original list
+        for index, desc in pending_line_items.items():
+            # Get result from API response, or None if API failed for that item
+            forto_code = api_results.get(desc)
+
+            # Update the original item
+            line_items[index]["itemCode"] = {
+                "documentValue": desc,
+                "formattedValue": forto_code,  # Might be None if API failed
+                "page": line_items[index]["lineItemDescription"].get("page"),
+            }
 
-def compute_score(args):
-    """Compute the fuzzy matching score between a new line item and a key."""
-    new_lineitem, key = args
-    return key, fuzz.ratio(new_lineitem, key)
+    # Add reverse charge here if exists
+    if reverse_charge:
+        [
+            item.update({"reverseChargeSentence": reverse_charge})
+            for item in line_items
+            if item["itemCode"]["formattedValue"] != "CDU"
+        ]
+    return line_items
 
 
 def get_fuzzy_match_score(target: str, sentences: list, threshold: int):
@@ -195,16 +246,18 @@ def get_fuzzy_match_score(target: str, sentences: list, threshold: int):
         tuple: (best_match, score) if above threshold, else (None, 0)
     """
     # Use multiprocessing to find the best match
-    with ThreadPoolExecutor() as executor:
-        results = executor.map(compute_score, [(target, s) for s in sentences])
+    result = process.extractOne(
+        target, sentences, scorer=fuzz.WRatio, score_cutoff=threshold
+    )
 
-    # Find the best match and score
-    best_match, best_score = max(results, key=lambda x: x[1], default=(None, 0))
+    if result is None:
+        return None, False
 
-    # return best_match, best_score
-    # If the best match score is above a threshold (e.g., 80), return it
-    if best_score >= threshold:
-        return best_match, True
+    match, score, index = result
+
+    # return best_match if the best match score is above a threshold (e.g., 80)
+    if match:
+        return match, True
 
     return None, False
 
@@ -236,46 +289,59 @@ def find_matching_lineitem(new_lineitem: str, kvp_dict: dict, threshold=90):
     Returns:
         str: The best matching 'Forto SLI' value from the dictionary.
     """
-    new_lineitem = new_lineitem.upper()
-
     # Check if the new line item is already in the dictionary
     if new_lineitem in kvp_dict:
         return kvp_dict[new_lineitem]
 
     # Get the best fuzzy match score for the extracted line item
-    best_match, _ = get_fuzzy_match_score(
-        new_lineitem, list(kvp_dict.keys()), threshold
+    match, _ = get_fuzzy_match_score(
+        new_lineitem,
+        list(kvp_dict.keys()),
+        threshold,
     )
 
-    return kvp_dict.get(best_match, None)
+    if match:
+        # find the code from the kvp_dict
+        return kvp_dict[match]
 
+    return None
 
-def associate_forto_item_code(input_string, params):
-    """
-    Finds a match for the input string using fuzzy matching first, then embedding fallback.
-
-    1. Tries to find a fuzzy match for input_string against the keys in
-       mapping_data using RapidFuzz, requiring a score >= fuzzy_threshold.
-    2. If found, returns the corresponding value from mapping_data.
-    3. If not found above threshold, calls the embedding_fallback function.
 
+async def associate_forto_item_code(line_item_data, params):
+    """
+    Associates Forto item codes to a list of line item descriptions.
     Args:
-        input_string: The string to find a match for.
-        params: Parameters containing the lookup data and fuzzy threshold.
+        line_item_data (dict): A dictionary where keys are original descriptions and values are cleaned descriptions.
+        params (dict): Parameters containing lookup data and thresholds.
 
     Returns:
-        The matched value (from fuzzy match or embedding), or None if no match found.
+        list: A list of dictionaries with 'description' and 'itemCode' keys.
     """
-    # Get the Forto item code using fuzzy matching
-    forto_item_code = find_matching_lineitem(
-        new_lineitem=input_string,
-        kvp_dict=params["lookup_data"]["item_code"],  # TODO: Parse the KVP dictionary
-        threshold=params["fuzzy_threshold_item_code"],
-    )
 
-    if forto_item_code is None:
-        # 2. Fallback to embedding function if no good fuzzy match
-        forto_item_code = get_tms_mappings(input_string, "line_items")
+    result = []
+    pending_line_items = {}
+    for desc, f_desc in line_item_data.items():
+        # Get the Forto item code using fuzzy matching
+        code = find_matching_lineitem(
+            new_lineitem=f_desc,
+            kvp_dict=params["lookup_data"]["item_code"],
+            threshold=params["fuzzy_threshold_item_code"],
+        )
+        if code:
+            result.append({"description": desc, "itemCode": code})
+        else:
+            pending_line_items[desc] = f_desc
+
+    # Batch API Call for Embedding lookups
+    if pending_line_items:
+        api_results = await get_tms_mappings(
+            input_list=list(pending_line_items.values()),
+            embedding_type="line_items",
+        )
+
+        # Merge API results back into original list
+        for desc, f_desc in pending_line_items.items():
+            code = api_results.get(f_desc)
+            result.append({"description": desc, "itemCode": code})
 
-    result = {"documentValue": input_string, "formattedValue": forto_item_code}
     return result

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/prompts/library/customsInvoice/other/prompt.txt

@@ -54,7 +54,7 @@ Your role is to accurately extract specific entities from these invoices to supp
     - unitPrice: Even if the quantity is not mentioned, you can still extract the unit price. Check the naming of the columns in a different languages, it can be "Unit Price", "Prezzo unitario", "Prix Unitaire", "Unitario", etc. Refer to "Prezzo unitario" field in the italian invoice example.
     - totalAmount: The total amount for the item. It can be in different currencies, so ensure to capture the currency as well for the totalAmountCurrency.
     - totalAmountEuro: Few line items contains a total amount in Euro. You can find it by looking for the term "Total EUR" or "Amount in Euro" in the line item but it's always in the EURO / € currency. Sometimes, it can be same as totalAmount if the line item is already in Euro.
-    - quantity: The quantity of the item or service provided in the line item.
+    - quantity: The quantity of the item or service provided in the line item. Pay attention to 2 x 40HC or 2x40HC. It means, quantity is 2 and 40HC is containerSize but not 240.
     - containerNumber: Container Number always starts with 4 letters and is followed by 7 digits (e.g., ABCD1234567).
 
 - hblNumber and mblNumber:

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/prompts/library/partnerInvoice/other/prompt.txt

@@ -52,8 +52,8 @@ Your role is to accurately extract specific entities from these invoices to supp
     - unitPrice: Even if the quantity is not mentioned, you can still extract the unit price. Check the naming of the columns in a different languages, it can be "Unit Price", "Prezzo unitario", "Prix Unitaire", "Unitario", etc. Refer to "Prezzo unitario" field in the italian invoice example.
     - totalAmount: The total amount for the item. It can be in different currencies, so ensure to capture the currency as well for the totalAmountCurrency.
     - totalAmountEuro: Few line items contains a total amount in Euro. You can find it by looking for the term "Total EUR" or "Amount in Euro" in the line item but it's always in the EURO / € currency. Sometimes, it can be same as totalAmount if the line item is already in Euro.
-    - quantity: The quantity of the item or service provided in the line item. Pay attention to 2x40HC. It means, quantity is 2 and containerSize is 40HC but not 240.
-    - containerNumber: Container Number always starts with 4 letters and is followed by 7 digits (e.g., ABCD1234567).
+    - quantity: The quantity of the item or service provided in the line item. Pay attention to 2 x 40HC or 2x40HC. It means, quantity is 2 and 40HC is containerSize but not 240.
+    - containerNumber: Container Number always starts with 4 letters and is followed by 7 digits (e.g., ABCD1234567, XALU 8593678).
 
 - hblNumber and mblNumber:
     - The Master Bill of Lading number. Commonly known as "Bill of Lading Number", "BILL OF LADING NO.", "BL Number", "BL No.", "B/L No.", "BL-Nr.", "B/L", or "HBL No.".
@@ -68,6 +68,7 @@ Your role is to accurately extract specific entities from these invoices to supp
     - Example:
         - "COSCO SHIPPING Lines Italy, Poland, or France S.R.L. – Genova Office – As Agent For COSCO SHIPPING Lines Co.,Ltd."
         - vendorName: COSCO SHIPPING Lines Co.,Ltd.
+    - From Hapag-Lloyd invoices, look for "Ballindamm 25" address to extract the vendorAddress.
 
 - agentName: Name of the agent. Agencies are offices authorized to act on behalf of a company. This details usually available including the branch name of the parent company name in the invoice.
 - agentKeyWord:
@@ -81,6 +82,7 @@ Your role is to accurately extract specific entities from these invoices to supp
 
 IMPORTANT NOTE:
 - Ensure all extracted values are directly from the document. Do not make assumptions, modifications or calculations.
+- Do not split the quantity into different line items. e.g., if quantity is 2 or 2 CTR or 2 BIL, do not create 2 separate line items with quantity 1 each.
 - Do not normalize or modify any entity values.
 - Pay attention to the line item details and paymentInformation, as they may vary significantly across different invoices.
 

{data_science_document_ai-1.43.6 → data_science_document_ai-1.45.2}/src/utils.py

@@ -6,16 +6,16 @@ import json
 import os
 import pickle
 from datetime import datetime
-from typing import Literal
+from typing import Any, Dict, List, Literal, Optional
 
+import httpx
 import numpy as np
 import openpyxl
 import pandas as pd
-import requests
 from google.cloud import documentai_v1beta3 as docu_ai_beta
 from pypdf import PdfReader, PdfWriter
 
-from src.io import get_storage_client, logger
+from src.io import bq_logs, get_storage_client, logger
 
 
 def get_pdf_page_count(pdf_bytes):
@@ -31,29 +31,6 @@ def get_pdf_page_count(pdf_bytes):
     return len(reader.pages)
 
 
-def bq_logs(data_to_insert, params):
-    """Insert logs into Google BigQuery.
-
-    Args:
-        data_to_insert (list): The data to insert into BigQuery.
-        params (dict): The parameters dictionary.
-    """
-    # Use the pre-initialized BigQuery client
-    bq_client = params["bq_client"]
-    # Get the table string
-    table_string = f"{params['g_ai_project_name']}.{params['g_ai_gbq_db_schema']}.{params['g_ai_gbq_db_table_out']}"
-
-    logger.info(f"Log table: {table_string}")
-    # Insert the rows into the table
-    insert_logs = bq_client.insert_rows_json(table_string, data_to_insert)
-
-    # Check if there were any errors inserting the rows
-    if not insert_logs:
-        logger.info("New rows have been added.")
-    else:
-        logger.info("Errors occurred while inserting rows: ", insert_logs)
-
-
 async def get_data_set_schema_from_docai(
     schema_client, project_id=None, location=None, processor_id=None, name=None
 ):
@@ -383,9 +360,9 @@ def extract_top_pages(pdf_bytes, num_pages=4):
     return output.getvalue()
 
 
-def get_tms_mappings(
-    input_list: list[str], embedding_type: str, llm_ports: list[str] = None
-):
+async def get_tms_mappings(
+    input_list: List[str], embedding_type: str, llm_ports: Optional[List[str]] = None
+) -> Dict[str, Any]:
     """Get TMS mappings for the given values.
 
     Args:
@@ -395,39 +372,66 @@ def get_tms_mappings(
         llm_ports (list[str], optional): List of LLM ports to use. Defaults to None.
 
     Returns:
-        dict: A dictionary with the mapping results.
+        dict or string: A dictionary or a string with the mapping results.
     """
-    # To test the API locally, port-forward the embedding service in the sandbox to 8080:80
-    # If you want to launch uvicorn from the tms-embedding repo, then use --port 8080 in the config file
     base_url = (
         "http://0.0.0.0:8080/"
         if os.getenv("CLUSTER") is None
         else "http://tms-mappings.api.svc.cluster.local./"
     )
 
+    # Ensure clean inputs
+    if not input_list:
+        return {}
+
     # Ensure input_list is a list
     if not isinstance(input_list, list):
         input_list = [input_list]
 
     # Always send a dict with named keys
     payload = {embedding_type: input_list}
+
     if llm_ports:
         payload["llm_ports"] = llm_ports if isinstance(llm_ports, list) else [llm_ports]
 
     # Make the POST request to the TMS mappings API
-    url = f"{base_url}/{embedding_type}"
-    response = requests.post(url=url, json=payload)
+    url = f"{base_url}{embedding_type}"
 
-    if response.status_code != 200:
-        logger.error(
-            f"Error from TMS mappings API: {response.status_code} - {response.text}"
-        )
+    # Use a timeout so the code doesn't hang forever
+    timeout = httpx.Timeout(60.0, connect=10.0)
+
+    async with httpx.AsyncClient(timeout=timeout) as client:
+        try:
+            response = await client.post(url, json=payload)
+            response.raise_for_status()
 
-    formatted_values = (
-        response.json().get("response", {}).get("data", {}).get(input_list[0], None)
+            # Structure expected: {"response": {"data": {"desc1": "code1", "desc2": "code2"}}}
+            return response.json().get("response", {}).get("data", {})
+
+        except httpx.HTTPStatusError as exc:
+            logger.error(
+                f"Error from TMS mappings API: {exc.response.status_code} - {exc.response.text}"
+            )
+            return {}
+
+
+async def batch_fetch_all_mappings(container_types, terminals, depots):
+    """Batch fetch all mappings for container types, terminals, and depots."""
+    # run batch calls concurrently
+    results = await asyncio.gather(
+        get_tms_mappings(list(container_types), "container_types"),
+        get_tms_mappings(list(terminals), "terminals"),
+        get_tms_mappings(list(depots), "depots"),
     )
 
-    return formatted_values
+    batch_container_map, batch_terminal_map, batch_depot_map = results
+
+    # Convert lists of tuples to dicts if necessary
+    return (
+        dict(batch_container_map or {}),
+        dict(batch_terminal_map or {}),
+        dict(batch_depot_map or {}),
+    )
 
 
 def transform_schema_strings(schema):
@@ -502,3 +506,21 @@ def estimate_page_count(sheet):
     else:
         return None
     return np.ceil(pg_cnt / 500)
+
+
+def split_pdf_into_chunks(file_content: bytes, chunk_size: int = 1):
+    """Split PDF into smaller page chunks."""
+    pdf = PdfReader(io.BytesIO(file_content))
+    total_pages = len(pdf.pages)
+
+    # TODO: update the chunk_size based on doc length. However, it breaks the page number extraction logic.
+    for i in range(0, total_pages, chunk_size):
+        writer = PdfWriter()
+        for j in range(i, min(i + chunk_size, total_pages)):
+            writer.add_page(pdf.pages[j])
+
+        buffer = io.BytesIO()
+        writer.write(buffer)
+        buffer.seek(0)
+
+        yield buffer.getvalue()