data-science-document-ai 1.37.0__py3-none-any.whl → 1.51.0__py3-none-any.whl

src/postprocessing/common.py

@@ -1,426 +1,22 @@
 import asyncio
-import datetime
 import json
 import os
 import re
 from datetime import timezone
 
-import numpy as np
 import pandas as pd
-import requests
 from nltk.corpus import stopwords
 from rapidfuzz import process
-from vertexai.preview.language_models import TextEmbeddingModel
 
 from src.constants import formatting_rules
-from src.io import get_storage_client, logger
+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.tms import call_tms, set_tms_service_token
+from src.utils import batch_fetch_all_mappings, get_tms_mappings
 
 tms_domain = os.environ["TMS_DOMAIN"]
 
 
-class EmbeddingsManager:  # noqa: D101
-    def __init__(self, params):  # noqa: D107
-        self.params = params
-        self.embeddings_dict = {}
-        self.embed_model = setup_embed_model()
-        self.bucket = self.get_bucket_storage()
-        self.embedding_folder = self.embed_model._model_id
-        self.embedding_dimension = 768  # TODO: to be reduced
-
-    def get_bucket_storage(self):
-        """
-        Retrieve the bucket storage object.
-
-        Returns:
-            The bucket storage object.
-        """
-        params = self.params
-        storage_client = get_storage_client(params)
-        bucket = storage_client.bucket(params["doc_ai_bucket_name"])
-        return bucket
-
-    def _find_most_similar_option(self, input_string, option_ids, option_embeddings):
-        """
-        Find the most similar option to the given input string based on embeddings.
-
-        Args:
-            model: The model used for generating embeddings.
-            input_string (str): The input string to find the most similar option for.
-            option_ids (list): The list of option IDs.
-            option_embeddings (np.ndarray): The embeddings of the options.
-
-        Returns:
-            The ID of the most similar option.
-        """
-        try:
-            input_embedding = self.embed_model.get_embeddings(
-                [input_string], output_dimensionality=self.embedding_dimension
-            )[0].values
-            similarities = np.dot(option_embeddings, input_embedding)
-            idx = np.argmax(similarities)
-            return option_ids[idx]
-        except Exception as e:
-            logger.error(f"Embeddings error: {e}")
-            return None
-
-    def load_embeddings(self):
-        """
-        Load embeddings for container types, ports, and terminals.
-
-        Returns:
-            None
-        """
-        for data_field in [
-            "container_types",
-            "ports",
-            "terminals",
-            "depots",
-            "item_codes_label",
-        ]:
-            self.embeddings_dict[data_field] = load_embed_by_data_field(
-                self.bucket,
-                f"{self.embedding_folder}/{data_field}/output",
-                self.embedding_dimension,
-            )
-
-    async def update_embeddings(self):
-        """
-        Update the embeddings dictionary.
-
-        Returns:
-            dict: The updated embeddings dictionary with the following keys:
-                - "container_types": A tuple containing the container types and their embeddings.
-                - "ports": A tuple containing the ports and their embeddings.
-                - "terminals": A tuple containing the terminal IDs and their embeddings.
-        """
-        # Update embeddings dict here.
-        # Ensure this method is async if you're calling async operations.
-        set_tms_service_token()
-        (
-            container_types,
-            container_type_embeddings,
-        ) = self.setup_container_type_embeddings(
-            *self.embeddings_dict.get("container_types", ([], []))
-        )
-
-        ports, port_embeddings = self.setup_ports_embeddings(
-            *self.embeddings_dict.get("ports", ([], []))
-        )
-
-        # Setup terminal embeddings
-        # Since retrieving terminal attributes requires calling TMS' api to extract terminals by each port,
-        # we only do it for new ports.
-        prev_port_ids, _ = self.embeddings_dict.get("ports", ([], []))
-        added_port_ids = [port for port in ports if port not in prev_port_ids]
-        if added_port_ids:
-            terminal_ids, terminal_embeddings = self.setup_terminal_embeddings(
-                added_port_ids
-            )
-        else:
-            terminal_ids, terminal_embeddings = self.embeddings_dict["terminals"]
-
-        depot_names, depot_embeddings = self.setup_depot_embeddings(
-            *self.embeddings_dict.get("depots", ([], []))
-        )
-
-        item_code_names, item_code_embeddings = self.setup_item_code_embeddings(
-            *self.embeddings_dict.get("item_codes_label", ([], []))
-        )
-
-        self.embeddings_dict = {
-            "container_types": (container_types, container_type_embeddings),
-            "ports": (ports, port_embeddings),
-            "terminals": (terminal_ids, terminal_embeddings),
-            "depots": (depot_names, depot_embeddings),
-            "item_codes_label": (item_code_names, item_code_embeddings),
-        }
-        return self.embeddings_dict
-
-    def batch_embed(self, option_strings: list[dict], suffix: str):
-        """
-        Compute embeddings for a batch of option strings and uploads them to a cloud storage bucket.
-
-        Args:
-            option_strings (list): A list of option strings to compute embeddings for.
-            suffix (str): A suffix to be used in the storage path for the embeddings:
-            input & output will be stored under "{bucket}/{parent_folder}/{suffix}/"
-
-        Returns:
-            tuple: A tuple containing the option IDs and embeddings.
-        """
-        now = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
-        input_path = f"{self.embedding_folder}/{suffix}/input/{now}.jsonl"
-        blob = self.bucket.blob(input_path)
-
-        # Convert each dictionary to a JSON string and join them with newlines
-        option_strings = [
-            {**option, "task_type": "SEMANTIC_SIMILARITY", "output_dimensionality": 256}
-            for option in option_strings
-        ]
-        jsonl_string = "\n".join(json.dumps(d) for d in option_strings)
-
-        # Convert the combined string to bytes
-        jsonl_bytes = jsonl_string.encode("utf-8")
-
-        # Upload the bytes to the blob
-        blob.upload_from_string(jsonl_bytes, content_type="text/plain")
-
-        # Compute embeddings for the options
-        embedding_path = f"{self.embedding_folder}/{suffix}/output"
-        assert len(option_strings) <= 30000  # Limit for batch embedding
-        batch_resp = self.embed_model.batch_predict(
-            dataset=f"gs://{self.bucket.name}/{input_path}",  # noqa
-            destination_uri_prefix=f"gs://{self.bucket.name}/{embedding_path}",  # noqa
-        )
-
-        if batch_resp.state.name != "JOB_STATE_SUCCEEDED":
-            logger.warning(
-                f"Batch prediction job failed with state {batch_resp.state.name}"
-            )
-        else:
-            logger.info(f"Embeddings for {suffix} computed successfully.")
-
-        option_ids, option_embeddings = load_embed_by_data_field(
-            self.bucket, embedding_path, self.embedding_dimension
-        )
-        return option_ids, option_embeddings
-
-    def setup_container_type_embeddings(
-        self, computed_container_type_ids, computed_container_type_embeddings
-    ):
-        """
-        Set up container type embeddings.
-
-        Args:
-            computed_container_type_ids (list): The list of already computed container type IDs.
-            computed_container_type_embeddings (list): The list of already computed container type embeddings.
-
-        Returns:
-            tuple: A tuple containing the updated container type IDs and embeddings.
-        """
-        url = (
-            f"https://tms.forto.{tms_domain}/api/transport-units/api/types/list"  # noqa
-        )
-        resp = call_tms(requests.get, url)
-        container_types = resp.json()
-
-        container_attribute_strings = [
-            dict(
-                title=container_type["code"],
-                content=" | ".join(
-                    [container_type["code"]]
-                    + [
-                        f"{v}"
-                        for k, v in container_type["containerAttributes"].items()
-                        if k in ["isoSizeType", "isoTypeGroup", "containerCategory"]
-                    ]
-                    + [container_type.get(k, "") for k in ["displayName", "notes"]]
-                ),
-            )
-            for container_type in container_types
-            if container_type["type"] == "Container"
-            and container_type["code"] not in computed_container_type_ids
-            and container_type.get("containerAttributes") is not None
-        ]
-        if not container_attribute_strings:
-            logger.info("No new container types found.")
-            return computed_container_type_ids, computed_container_type_embeddings
-
-        logger.info("Computing embeddings for container types...")
-        container_type_ids, container_type_embeddings = self.batch_embed(
-            container_attribute_strings, "container_types"
-        )
-        return container_type_ids, container_type_embeddings
-
-    def setup_ports_embeddings(self, computed_port_ids, computed_port_embeddings):
-        """
-        Set up port embeddings.
-
-        Steps:
-        - Retrieve active ports from the TMS API
-        - Compute embeddings for new tradelane-enabled ports
-        - Return ALL port IDs and embeddings.
-
-        Args:
-            computed_port_ids (list): The list of previously computed port IDs.
-            computed_port_embeddings (list): The list of previously computed port embeddings.
-
-        Returns:
-            tuple: A tuple containing ALL port IDs and embeddings.
-        """
-        url = f"https://tms.forto.{tms_domain}/api/transport-network/api/ports?pageSize=1000000&status=active"  # noqa
-        resp = call_tms(requests.get, url)
-        resp_json = resp.json()
-        if len(resp_json["data"]) != resp_json["_paging"]["totalRecords"]:
-            logger.error("Not all ports were returned.")
-
-        new_sea_ports = [
-            port
-            for port in resp_json["data"]
-            if "sea" in port["modes"] and port["id"] not in computed_port_ids
-        ]
-        if not new_sea_ports:
-            logger.info("No new ports found.")
-            return computed_port_ids, computed_port_embeddings
-
-        port_attribute_strings = [
-            dict(
-                title=port["id"],
-                content=" ".join(
-                    [
-                        "port for shipping",
-                        add_text_without_space(
-                            port["name"]
-                        ),  # for cases like QUINHON - Quinhon
-                        port["id"],
-                    ]
-                ),
-            )
-            for port in new_sea_ports
-        ]
-
-        logger.info("Computing embeddings for ports.")
-        port_ids, port_embeddings = self.batch_embed(port_attribute_strings, "ports")
-        return port_ids, port_embeddings
-
-    def setup_depot_embeddings(self, computed_depot_names, computed_depot_embeddings):
-        """
-        Set up depot embeddings.
-
-        Steps:
-        - Retrieve active depot from the TMS API
-        - Compute embeddings for new tdepot
-        - Return ALL depot names and embeddings.
-
-        Args:
-            computed_depot_names (list): The list of previously computed depot names.
-            computed_depot_embeddings (list): The list of previously computed depot embeddings.
-
-        Returns:
-            tuple: A tuple containing ALL depot names and embeddings.
-        """
-        url = f"https://tms.forto.{tms_domain}/api/transport-network/api/depots?pageSize=1000000"  # noqa
-        resp = call_tms(requests.get, url)
-        resp_json = resp.json()
-
-        new_depots = [
-            depot
-            for depot in resp_json["data"]
-            if depot["name"] not in computed_depot_names
-        ]
-        if not new_depots:
-            logger.info("No new depots found.")
-            return computed_depot_names, computed_depot_embeddings
-
-        depot_attribute_strings = [
-            dict(
-                title=depot["name"],
-                content=" | ".join(
-                    [
-                        "depot",
-                        "name - " + depot["name"],
-                        "address - " + depot["address"]["fullAddress"],
-                    ]
-                ),
-            )
-            for depot in resp_json["data"]
-        ]
-
-        logger.info("Computing embeddings for depots.")
-        depot_names, depot_embeddings = self.batch_embed(
-            depot_attribute_strings, "depots"
-        )
-        return depot_names, depot_embeddings
-
-    def setup_terminal_embeddings(self, added_port_ids):
-        """
-        Set up terminal embeddings for `added_port_ids`, using `model`, uploaded to `bucket`.
-
-        Args:
-            added_port_ids (list): A list of added port IDs.
-
-        Returns:
-            tuple: A tuple containing the ALL terminal IDs and terminal embeddings.
-            Not just for the added port IDs.
-        """
-        terminal_attibute_strings = [
-            setup_terminal_attributes(port_id) for port_id in added_port_ids
-        ]
-        terminal_attibute_strings = sum(terminal_attibute_strings, [])
-        if not terminal_attibute_strings:
-            logger.info("No new terminals found.")
-            return [], np.array([])
-
-        terminal_ids, terminal_embeddings = self.batch_embed(
-            terminal_attibute_strings, "terminals"
-        )
-        return terminal_ids, terminal_embeddings
-
-    def setup_item_code_embeddings(
-        self, computed_item_code_names, computed_item_code_embeddings
-    ):
-        """
-        Set up item_code embeddings.
-
-        Steps:
-        - Retrieve active item_code from the TMS API
-        - Compute embeddings for new titem_code
-        - Return ALL item_code names and embeddings.
-
-        Args:
-            computed_item_code_names (list): The list of previously computed item_code names.
-            computed_item_code_embeddings (list): The list of previously computed item_code embeddings.
-
-        Returns:
-            tuple: A tuple containing ALL item_code names and embeddings.
-        """
-        url = f"https://tms.forto.{tms_domain}/api/catalog/item-codes?transportTypes=fcl&pageSize=1000000"  # noqa
-        resp = call_tms(requests.get, url)
-        resp_json = resp.json()
-
-        new_item_codes = [
-            item_code
-            for item_code in resp_json["results"]
-            if item_code["id"] not in computed_item_code_names
-        ]
-        if not new_item_codes:
-            logger.info("No new item_codes found.")
-            return computed_item_code_names, computed_item_code_embeddings
-
-        item_code_attribute_strings = [
-            dict(
-                title=item_code["id"],
-                content=" | ".join(
-                    [
-                        item_code["id"],
-                        item_code["label"],
-                    ]
-                ),
-            )
-            for item_code in resp_json["results"]
-        ]
-
-        logger.info("Computing embeddings for item_codes.")
-        item_code_names, item_code_embeddings = self.batch_embed(
-            item_code_attribute_strings, "item_codes_label"
-        )
-        return item_code_names, item_code_embeddings
-
-
-def setup_embed_model():
-    """
-    Set up and return a text embedding model.
-
-    Returns:
-        TextEmbeddingModel: The initialized text embedding model.
-    """
-    model = TextEmbeddingModel.from_pretrained("text-multilingual-embedding-002")
-    return model
-
-
 def convert_container_number(container_number):
     """
     Convert a container number to ISO standard.
@@ -488,16 +84,16 @@ def clean_shipment_id(shipment_id):
     """
     if not shipment_id:
         return
-    # '#S123456@-1' -> 'S123456'
-    # Find the pattern of a shipment ID that starts with 'S' followed by 5 to 7 digits
-    match = re.findall(r"S\d{5,7}", shipment_id)
+    # '#S1234565@-1' -> 'S1234565'
+    # Find the pattern of a shipment ID that starts with 'S' followed by 7 to 8 digits
+    match = re.findall(r"S\d{6,8}", shipment_id)
     stripped_value = match[0] if match else None
 
     if not stripped_value:
         return None
 
     # Check if length is valid (should be either 7 or 8)
-    if len(stripped_value) not in (6, 7, 8):
+    if len(stripped_value) not in (7, 8, 9):
         return None
 
     return stripped_value
@@ -538,9 +134,12 @@ 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:
-        if c.isnumeric() or c in [",", "."]:
+    for c in value:
+        if c.isnumeric() or c in [",", ".", "-"]:
             formatted_value += c
 
     # First and last characters should not be  [",", "."]
@@ -570,106 +169,6 @@ def extract_string(data_field_value):
     return formatted_value if formatted_value not in ["''", ""] else None
 
 
-def extract_google_embed_resp(prediction_string, embedding_dimension):
-    """
-    Extract relevant information from the Google Embed API response.
-
-    Args:
-        prediction_string (str): The prediction string returned by the Google Embed API.
-
-    Returns:
-        dict: A dictionary containing the extracted information.
-            - _id (str): The title of the instance.
-            - attr_text (str): The content of the instance.
-            - embedding (list): The embeddings values from the predictions.
-
-    """
-    res = json.loads(prediction_string)
-    return dict(
-        _id=res["instance"]["title"],
-        attr_text=res["instance"]["content"],
-        embedding=res["predictions"][0]["embeddings"]["values"][:embedding_dimension],
-    )
-
-
-def load_embed_by_data_field(bucket, embedding_path, embedding_dimension):
-    """
-    Load embeddings by data field from the specified bucket and embedding path.
-
-    Args:
-        bucket (Bucket): The bucket object representing the storage bucket.
-        embedding_path (str): The path to the embeddings in the bucket (different by data_field).
-
-    Returns:
-        tuple: A tuple containing the option IDs and option embeddings.
-            - option_ids (list): A list of option IDs.
-            - option_embeddings (ndarray): An array of option embeddings.
-    """
-    # Retrieve the embeddings from the output files
-    blobs = bucket.list_blobs(prefix=embedding_path)
-    all_blob_data = []
-    for blob in blobs:
-        blob_data = blob.download_as_bytes().decode("utf-8").splitlines()
-        embeddings = [
-            extract_google_embed_resp(data, embedding_dimension) for data in blob_data
-        ]
-        all_blob_data.extend(embeddings)
-    option_ids = [embed["_id"] for embed in all_blob_data]
-    option_embeddings = np.stack([embed["embedding"] for embed in all_blob_data])
-    return option_ids, option_embeddings
-
-
-def setup_terminal_attributes(port_id: str):
-    """
-    Retrieve and format the attributes of active terminals at a given port.
-
-    Args:
-        port_id (str): The ID of the port.
-
-    Returns:
-        list: A list of dictionaries containing the formatted attributes of active terminals.
-              Each dictionary has the following keys:
-              - title: The terminal's short code.
-              - content: A string representation of the terminal's attributes, including its name,
-                         searchable name, and full address.
-    """
-    url = f"https://gateway.forto.{tms_domain}/api/transport-network/api/ports/{port_id}/terminals/list"  # noqa
-    resp = call_tms(requests.get, url)
-    terminals = resp.json()
-    if len(terminals) == 0:
-        return []
-    active_terminals = [term for term in terminals if term["isActive"]]
-    if len(active_terminals) == 0:
-        logger.warning(f"No active terminals found at port {port_id}.")
-        return []
-
-    terminal_attibute_strings = [
-        dict(
-            title=term["name"],
-            content=" | ".join(
-                [
-                    "shipping terminal",
-                    "code - " + term["terminalShortCode"],
-                    "name - " + modify_terminal_name(term["searchableName"]),
-                    "country - " + term["address"]["country"],
-                ]
-            ),
-        )
-        for term in active_terminals
-    ]
-    return terminal_attibute_strings
-
-
-def modify_terminal_name(text):
-    # Find the first occurrence of a word starting with 'K' followed by a number
-    # and replace it with 'KAAI' - meaning Quay in Dutch
-    match = re.search(r"K(\d+)", text)
-    if match:
-        # Append "KAAI" followed by the number if a match is found
-        text += f" KAAI {match.group(1)}"
-    return text
-
-
 def remove_none_values(d):
     if isinstance(d, dict):
         # Create a new dictionary to store non-None values
@@ -731,25 +230,6 @@ def convert_invoice_type(data_field_value, params):
     return None
 
 
-def validate_reverse_charge_value(reverse_charge_sentence_value):
-    """
-    Validates the reverseChargeSentence value before assigning to line items.
-
-    Args:
-        reverse_charge_sentence_value (bool): The formatted value of reverseChargeSentence (True or False).
-
-    Returns:
-        bool: The validated reverseChargeSentence value.
-    """
-    if isinstance(reverse_charge_sentence_value, bool):
-        return reverse_charge_sentence_value
-    else:
-        logger.warning(
-            f"Invalid reverseChargeSentence value: {reverse_charge_sentence_value}. Defaulting to False."
-        )
-        return False
-
-
 # Function to create KVP dictionary using apply method
 def create_kvp_dictionary(df_raw: pd.DataFrame):
     """Create a key-value pair dictionary from the given DataFrame.
@@ -842,6 +322,14 @@ def remove_unwanted_patterns(lineitem: str):
     # Remove "HIGH CUBE"
     lineitem = lineitem.replace("HIGH CUBE", "")
 
+    # Remove container size e.g., 20FT, 40HC, etc.
+    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
 
 
@@ -872,62 +360,92 @@ def clean_item_description(lineitem: str, remove_numbers: bool = True):
     # Remove the currency codes
     lineitem = re.sub(currency_codes_pattern, "", lineitem, flags=re.IGNORECASE)
 
+    # remove other patterns
+    lineitem = remove_unwanted_patterns(lineitem)
+
     # Remove numbers from the line item
     if (
         remove_numbers
     ):  # Do not remove numbers for the reverse charge sentence as it contains Article number
         lineitem = re.sub(r"\d+", "", lineitem)
 
-    # remove other patterns
-    lineitem = remove_unwanted_patterns(lineitem)
-
     # remove special chars
     lineitem = re.sub(r"[^A-Za-z0-9\s]", " ", lineitem).strip()
 
+    # Remove x from lineitem like 10 x
+    lineitem = re.sub(r"\b[xX]\b", " ", lineitem).strip()
+
     return re.sub(r"\s{2,}", " ", lineitem).strip()
 
 
 async def format_label(
-    entity_k, entity_value, embed_manager, document_type_code, params
+    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, embed_manager, document_type_code, params)
+            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, embed_manager, document_type_code, params)
+                format_label(
+                    entity_k,
+                    sub_v,
+                    document_type_code,
+                    params,
+                    mime_type,
+                    container_map,
+                    terminal_map,
+                    depot_map,
+                )
                 for sub_v in entity_value
             ]
         )
         return entity_k, [v for _, v in format_tasks]
+
+    if mime_type == "application/pdf":
+        if isinstance(entity_value, tuple):
+            page = entity_value[1]
+            entity_value = entity_value[0]
+        else:
+            page = -1
+
     entity_key = entity_k.lower()
-    embeddings_dict = embed_manager.embeddings_dict
     formatted_value = None
 
     if entity_key.startswith("port"):
         formatted_value = await get_port_code_ai(
-            entity_value, llm_client, embed_manager, *embeddings_dict["ports"]
+            entity_value, llm_client, doc_type=document_type_code
         )
+
     elif (entity_key == "containertype") or (entity_key == "containersize"):
-        formatted_value = embed_manager._find_most_similar_option(
-            "container type " + entity_value,
-            *embeddings_dict["container_types"],
-        )
+        formatted_value = container_map.get(entity_value)
+
     elif check_formatting_rule(entity_k, document_type_code, "terminal"):
-        formatted_value = embed_manager._find_most_similar_option(
-            "shipping terminal " + str(entity_value),
-            *embeddings_dict["terminals"],
-        )
+        formatted_value = terminal_map.get(entity_value)
+
     elif check_formatting_rule(entity_k, document_type_code, "depot"):
-        formatted_value = embed_manager._find_most_similar_option(
-            "depot " + str(entity_value),
-            *embeddings_dict["depots"],
-        )
+        formatted_value = depot_map.get(entity_value)
+
     elif entity_key.startswith(("eta", "etd", "duedate", "issuedate", "servicedate")):
         try:
             cleaned_data_field_value = clean_date_string(entity_value)
@@ -947,21 +465,26 @@ async def format_label(
         except ValueError as e:
             logger.info(f"ParserError: {e}")
 
-    elif entity_key in ["invoicenumber", "creditnoteinvoicenumber"]:
+    elif (
+        entity_key in ["invoicenumber", "creditnoteinvoicenumber"]
+        and document_type_code == "bundeskasse"
+    ):
         formatted_value = clean_invoice_number(entity_value)
 
     elif entity_key in ("shipmentid", "partnerreference"):
-        # Clean the shipment ID to match Forto's standard (starts with 'S' followed by 5 to 7 digits)
+        # Clean the shipment ID to match Forto's standard (starts with 'S' followed by 7 or 8 digits)
         formatted_value = clean_shipment_id(entity_value)
 
     elif entity_key == "containernumber":
         # Remove all non-alphanumeric characters like ' ', '-', etc.
         formatted_value = convert_container_number(entity_value)
 
-    elif (
-        document_type_code in ["finalMbL", "draftMbl"] and entity_key == "measurements"
+    elif any(
+        numeric_indicator in entity_key
+        for numeric_indicator in ["measurements", "weight"]
     ):
-        formatted_value = decimal_convertor(extract_number(entity_value))
+        formatted_value = extract_number(entity_value)
+
     elif any(
         packaging_type in entity_key
         for packaging_type in ["packagingtype", "packagetype", "currency"]
@@ -977,11 +500,19 @@ async def format_label(
     elif "reversechargesentence" in entity_key:
         formatted_value = clean_item_description(entity_value, remove_numbers=False)
 
+    elif "quantity" in entity_key:
+        if document_type_code in ["partnerInvoice", "customsInvoice", "bundeskasse"]:
+            # For partner invoice, quantity can be mentioned as whole number
+            # Apply decimal convertor for 46,45 --> 46.45 but not for 1.000 --> 1000
+            formatted_value = decimal_convertor(
+                extract_number(entity_value), quantity=True
+            )
+        else:
+            formatted_value = extract_number(entity_value)
+
     elif any(
         numeric_indicator in entity_key
         for numeric_indicator in [
-            "weight",
-            "quantity",
             "value",
             "amount",
             "price",
@@ -999,21 +530,21 @@ async def format_label(
         "documentValue": entity_value,
         "formattedValue": formatted_value,
     }
+    if mime_type == "application/pdf":
+        result["page"] = page
+
     return entity_k, result
 
 
-async def get_port_code_ai(
-    port: str, llm_client, embed_manager, port_ids, port_embeddings
-):
-    port_llm = await get_port_code_llm(port, llm_client)
+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)
 
-    if port_llm in port_ids:
-        return port_llm
-    port_text = f"port for shipping {port}"
-    return embed_manager._find_most_similar_option(port_text, port_ids, port_embeddings)
+    result = await get_tms_mappings(port, "ports", port_llm)
+    return result.get(port, None)
 
 
-async def get_port_code_llm(port: str, llm_client):
+async def get_port_code_llm(port: str, llm_client, doc_type=None):
     if (
         "postprocessing" in prompt_library.library.keys()
         and "port_code" in prompt_library.library["postprocessing"].keys()
@@ -1040,7 +571,7 @@ async def get_port_code_llm(port: str, llm_client):
         }
 
         response = await llm_client.get_unified_json_genai(
-            prompt, response_schema=response_schema, model="chatgpt"
+            prompt, response_schema=response_schema, model="chatgpt", doc_type=doc_type
         )
         try:
             mapped_port = response["port"]
@@ -1050,7 +581,7 @@ async def get_port_code_llm(port: str, llm_client):
             return None
 
 
-def decimal_convertor(value):
+def decimal_convertor(value, quantity=False):
     """Convert EU values to English values."""
     if value is None:
         return None
@@ -1058,30 +589,118 @@ def decimal_convertor(value):
     # Remove spaces
     value = value.strip().replace(" ", "")
 
-    # Convert comma to dot for decimal point (e.g., 4.123,45 -> 4123.45)
-    if re.match(r"^\d{1,3}(\.\d{3})*,\d{1,2}$", value):
-        value = value.replace(".", "").replace(",", ".")
+    # Check "-" and remove it for processing
+    is_negative, value = (True, value[1:]) if value.startswith("-") else (False, value)
+
+    if not quantity:
+        # Convert comma to dot for decimal point (e.g., 4.123,45 -> 4123.45)
+        if re.match(r"^\d{1,3}(\.\d{3})*,\d{1,2}$", value):
+            value = value.replace(".", "").replace(",", ".")
+
+        # European style integer with thousand separator: 2.500
+        elif re.match(r"^\d{1,3}(\.\d{3})+$", value):
+            value = value.replace(".", "")
+
+        # Format english values as well for consistency (e.g., 4,123.45 -> 4123.45)
+        elif re.match(r"^\d{1,3}(,\d{3})*\.\d{1,2}$", value):
+            value = value.replace(",", "")
 
-    # European style integer with thousand separator: 2.500
-    elif re.match(r"^\d{1,3}(\.\d{3})+$", value):
-        value = value.replace(".", "")
+        # English style integer with thousand separator: 2,500
+        elif re.match(r"^\d{1,3}(,\d{3})+$", value):
+            value = value.replace(",", "")
 
-    # Format english values as well for consistency (e.g., 4,123.45 -> 4123.45)
-    elif re.match(r"^\d{1,3}(,\d{3})*\.\d{1,2}$", value):
-        value = value.replace(",", "")
+        # Just replace comma decimals with dot (e.g., 65,45 -> 65.45)
+        if re.match(r"^\d+,\d{1,2}$", value):
+            value = value.replace(",", ".")
 
-    # English style integer with thousand separator: 2,500
-    elif re.match(r"^\d{1,3}(,\d{3})+$", value):
-        value = value.replace(",", "")
+        # If there are more than 3 0s after decimal point, consider only 2 decimal points (e.g., 8.500000 -> 8.50)
+        elif re.match(r"^\d+\.\d{3,}$", value):
+            value = value[: value.index(".") + 3]
 
-    # Just replace comma decimals with dot (e.g., 65,45 -> 65.45)
-    elif re.match(r"^\d+,\d{1,2}$", value):
-        value = value.replace(",", ".")
+    else:  # quantity=True → only last two
+        # Just replace comma decimals with dot (e.g., 65,45 -> 65.45)
+        if re.match(r"^\d+,\d{1,2}$", value):
+            value = value.replace(",", ".")
+
+        # If there are more than 3 0s after decimal point, consider only 2 decimal points (e.g., 8.500000 -> 8.50)
+        elif re.match(r"^\d+\.\d{3,}$", value):
+            value = value[: value.index(".") + 3]
+
+    # Re-add negative sign if applicable
+    value = "-" + value if is_negative else value
 
     return value
 
 
-async def format_all_entities(result, embed_manager, document_type_code, params):
+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`
     document_type_code = (
@@ -1096,15 +715,13 @@ async def format_all_entities(result, embed_manager, document_type_code, params)
         return {}
 
     # Format all entities recursively
-    _, aggregated_data = await format_label(
-        None, result, embed_manager, document_type_code, params
+    _, 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, embed_manager, document_type_code
-        )
+        await process_partner_invoice(params, aggregated_data, document_type_code)
 
     logger.info("Data Extraction completed successfully")
     return aggregated_data
@@ -1134,3 +751,48 @@ def remove_stop_words(lineitem: str):
         .upper()
         .strip()
     )
+
+
+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, 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, page_number=effective_page)
+                for v in llm_prediction
+            ]
+        else:
+            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"])
+                except:  # noqa: E722
+                    _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, 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, page_number
+            )
+
+    return llm_prediction