nucliadb 4.0.0.post542__py3-none-any.whl → 6.2.1.post2798__py3-none-any.whl

nucliadb/search/search/chat/prompt.py

@@ -17,32 +17,55 @@
 # You should have received a copy of the GNU Affero General Public License
 # along with this program. If not, see <http://www.gnu.org/licenses/>.
 #
+import asyncio
+import copy
+from collections import deque
 from dataclasses import dataclass
-from typing import Dict, List, Optional, Sequence, Tuple
+from typing import Deque, Dict, List, Optional, Sequence, Tuple, Union, cast
 
+import yaml
+from pydantic import BaseModel
+
+from nucliadb.common.ids import FIELD_TYPE_STR_TO_PB, FieldId, ParagraphId
+from nucliadb.common.maindb.utils import get_driver
+from nucliadb.common.models_utils import from_proto
 from nucliadb.ingest.fields.base import Field
 from nucliadb.ingest.fields.conversation import Conversation
+from nucliadb.ingest.fields.file import File
 from nucliadb.ingest.orm.knowledgebox import KnowledgeBox as KnowledgeBoxORM
-from nucliadb.ingest.orm.resource import KB_REVERSE
-from nucliadb.ingest.orm.resource import Resource as ResourceORM
-from nucliadb.middleware.transaction import get_read_only_transaction
 from nucliadb.search import logger
-from nucliadb.search.search import paragraphs
-from nucliadb.search.search.chat.images import get_page_image, get_paragraph_image
+from nucliadb.search.search import cache
+from nucliadb.search.search.chat.images import (
+    get_file_thumbnail_image,
+    get_page_image,
+    get_paragraph_image,
+)
+from nucliadb.search.search.hydrator import hydrate_field_text, hydrate_resource_text
+from nucliadb.search.search.paragraphs import get_paragraph_text
+from nucliadb_models.metadata import Extra, Origin
 from nucliadb_models.search import (
     SCORE_TYPE,
+    ConversationalStrategy,
+    FieldExtensionStrategy,
     FindParagraph,
+    FullResourceStrategy,
+    HierarchyResourceStrategy,
     ImageRagStrategy,
     ImageRagStrategyName,
-    KnowledgeboxFindResults,
+    MetadataExtensionStrategy,
+    MetadataExtensionType,
+    NeighbouringParagraphsStrategy,
+    PageImageStrategy,
+    ParagraphImageStrategy,
     PromptContext,
     PromptContextImages,
     PromptContextOrder,
     RagStrategy,
     RagStrategyName,
+    TableImageStrategy,
 )
 from nucliadb_protos import resources_pb2
-from nucliadb_utils.asyncio_utils import ConcurrentRunner, run_concurrently
+from nucliadb_utils.asyncio_utils import run_concurrently
 from nucliadb_utils.utilities import get_storage
 
 MAX_RESOURCE_TASKS = 5
@@ -53,6 +76,12 @@ MAX_RESOURCE_FIELD_TASKS = 4
 # The hope here is it will be enough to get the answer to the question.
 CONVERSATION_MESSAGE_CONTEXT_EXPANSION = 15
 
+TextBlockId = Union[ParagraphId, FieldId]
+
+
+class ParagraphIdNotFoundInExtractedMetadata(Exception):
+    pass
+
 
 class CappedPromptContext:
     """
@@ -70,16 +99,26 @@ class CappedPromptContext:
         self._size = 0
 
     def __setitem__(self, key: str, value: str) -> None:
+        prev_value_len = len(self.output.get(key, ""))
         if self.max_size is None:
-            # Unbounded size
-            self.output[key] = value
+            # Unbounded size context
+            to_add = value
         else:
-            existing_len = len(self.output.get(key, ""))
-            self._size -= existing_len
-            size_available = self.max_size - self._size
-            if size_available > 0:
-                self.output[key] = value[:size_available]
-                self._size += len(self.output[key])
+            # Make sure we don't exceed the max size
+            size_available = max(self.max_size - self._size + prev_value_len, 0)
+            to_add = value[:size_available]
+        self.output[key] = to_add
+        self._size = self._size - prev_value_len + len(to_add)
+
+    def __getitem__(self, key: str) -> str:
+        return self.output.__getitem__(key)
+
+    def __delitem__(self, key: str) -> None:
+        value = self.output.pop(key, "")
+        self._size -= len(value)
+
+    def text_block_ids(self) -> list[str]:
+        return list(self.output.keys())
 
     @property
     def size(self) -> int:
@@ -94,15 +133,15 @@ async def get_next_conversation_messages(
     num_messages: int,
     message_type: Optional[resources_pb2.Message.MessageType.ValueType] = None,
     msg_to: Optional[str] = None,
-):
+) -> List[resources_pb2.Message]:
     output = []
     cmetadata = await field_obj.get_metadata()
     for current_page in range(page, cmetadata.pages + 1):
         conv = await field_obj.db_get_value(current_page)
         for message in conv.messages[start_idx:]:
-            if message_type is not None and message.type != message_type:
+            if message_type is not None and message.type != message_type:  # pragma: no cover
                 continue
-            if msg_to is not None and msg_to not in message.to:
+            if msg_to is not None and msg_to not in message.to:  # pragma: no cover
                 continue
             output.append(message)
             if len(output) >= num_messages:
@@ -125,16 +164,21 @@ async def find_conversation_message(
 
 
 async def get_expanded_conversation_messages(
-    *, kb: KnowledgeBoxORM, rid: str, field_id: str, mident: str
+    *,
+    kb: KnowledgeBoxORM,
+    rid: str,
+    field_id: str,
+    mident: str,
+    max_messages: int = CONVERSATION_MESSAGE_CONTEXT_EXPANSION,
 ) -> list[resources_pb2.Message]:
     resource = await kb.get(rid)
-    if resource is None:
+    if resource is None:  # pragma: no cover
         return []
-    field_obj = await resource.get_field(field_id, KB_REVERSE["c"], load=True)
+    field_obj: Conversation = await resource.get_field(field_id, FIELD_TYPE_STR_TO_PB["c"], load=True)  # type: ignore
     found_message, found_page, found_idx = await find_conversation_message(
         field_obj=field_obj, mident=mident
     )
-    if found_message is None:
+    if found_message is None:  # pragma: no cover
         return []
     elif found_message.type == resources_pb2.Message.MessageType.QUESTION:
         # only try to get answer if it was a question
@@ -150,7 +194,7 @@ async def get_expanded_conversation_messages(
             field_obj=field_obj,
             page=found_page,
             start_idx=found_idx + 1,
-            num_messages=CONVERSATION_MESSAGE_CONTEXT_EXPANSION,
+            num_messages=max_messages,
         )
 
 
@@ -169,83 +213,27 @@ async def default_prompt_context(
     - Using an dict prevents from duplicates pulled in through conversation expansion.
     """
     # Sort retrieved paragraphs by decreasing order (most relevant first)
-    txn = await get_read_only_transaction()
-    storage = await get_storage()
-    kb = KnowledgeBoxORM(txn, storage, kbid)
-    for paragraph in ordered_paragraphs:
-        context[paragraph.id] = _clean_paragraph_text(paragraph)
-
-        # If the paragraph is a conversation and it matches semantically, we assume we
-        # have matched with the question, therefore try to include the answer to the
-        # context by pulling the next few messages of the conversation field
-        rid, field_type, field_id, mident = paragraph.id.split("/")[:4]
-        if field_type == "c" and paragraph.score_type in (
-            SCORE_TYPE.VECTOR,
-            SCORE_TYPE.BOTH,
-        ):
-            expanded_msgs = await get_expanded_conversation_messages(
-                kb=kb, rid=rid, field_id=field_id, mident=mident
-            )
-            for msg in expanded_msgs:
-                text = msg.content.text.strip()
-                pid = f"{rid}/{field_type}/{field_id}/{msg.ident}/0-{len(msg.content.text) + 1}"
-                context[pid] = text
-
-
-async def get_field_extracted_text(field: Field) -> Optional[tuple[Field, str]]:
-    extracted_text_pb = await field.get_extracted_text(force=True)
-    if extracted_text_pb is None:
-        return None
-    return field, extracted_text_pb.text
-
-
-async def get_resource_field_extracted_text(
-    kb_obj: KnowledgeBoxORM,
-    resource_uuid,
-    field_id: str,
-) -> Optional[tuple[Field, str]]:
-    resource = await kb_obj.get(resource_uuid)
-    if resource is None:
-        return None
-
-    try:
-        field_type, field_key = field_id.strip("/").split("/")
-    except ValueError:
-        logger.error(f"Invalid field id: {field_id}. Skipping getting extracted text.")
-        return None
-    field = await resource.get_field(field_key, KB_REVERSE[field_type], load=False)
-    if field is None:
-        return None
-    result = await get_field_extracted_text(field)
-    if result is None:
-        return None
-    _, extracted_text = result
-    return field, extracted_text
-
-
-async def get_resource_extracted_texts(
-    kbid: str,
-    resource_uuid: str,
-) -> list[tuple[Field, str]]:
-    txn = await get_read_only_transaction()
-    storage = await get_storage()
-    kb = KnowledgeBoxORM(txn, storage, kbid)
-    resource = ResourceORM(
-        txn=txn,
-        storage=storage,
-        kb=kb,
-        uuid=resource_uuid,
-    )
-
-    # Schedule the extraction of the text of each field in the resource
-    runner = ConcurrentRunner(max_tasks=MAX_RESOURCE_FIELD_TASKS)
-    for field_type, field_key in await resource.get_fields(force=True):
-        field = await resource.get_field(field_key, field_type, load=False)
-        runner.schedule(get_field_extracted_text(field))
-
-    # Wait for the results
-    results = await runner.wait()
-    return [result for result in results if result is not None]
+    async with get_driver().transaction(read_only=True) as txn:
+        storage = await get_storage()
+        kb = KnowledgeBoxORM(txn, storage, kbid)
+        for paragraph in ordered_paragraphs:
+            context[paragraph.id] = _clean_paragraph_text(paragraph)
+
+            # If the paragraph is a conversation and it matches semantically, we assume we
+            # have matched with the question, therefore try to include the answer to the
+            # context by pulling the next few messages of the conversation field
+            rid, field_type, field_id, mident = paragraph.id.split("/")[:4]
+            if field_type == "c" and paragraph.score_type in (
+                SCORE_TYPE.VECTOR,
+                SCORE_TYPE.BOTH,
+            ):
+                expanded_msgs = await get_expanded_conversation_messages(
+                    kb=kb, rid=rid, field_id=field_id, mident=mident
+                )
+                for msg in expanded_msgs:
+                    text = msg.content.text.strip()
+                    pid = f"{rid}/{field_type}/{field_id}/{msg.ident}/0-{len(msg.content.text) + 1}"
+                    context[pid] = text
 
 
 async def full_resource_prompt_context(
@@ -253,19 +241,18 @@ async def full_resource_prompt_context(
     kbid: str,
     ordered_paragraphs: list[FindParagraph],
     resource: Optional[str],
-    number_of_full_resources: Optional[int] = None,
+    strategy: FullResourceStrategy,
 ) -> None:
     """
     Algorithm steps:
         - Collect the list of resources in the results (in order of relevance).
         - For each resource, collect the extracted text from all its fields and craft the context.
-
     Arguments:
         context: The context to be updated.
         kbid: The knowledge box id.
-        results: The results of the retrieval (find) operation.
+        ordered_paragraphs: The results of the retrieval (find) operation.
         resource: The resource to be included in the context. This is used only when chatting with a specific resource with no retrieval.
-        number_of_full_resources: The number of full resources to include in the context.
+        strategy: strategy instance containing, for example, the number of full resources to include in the context.
     """  # noqa: E501
     if resource is not None:
         # The user has specified a resource to be included in the context.
@@ -274,32 +261,205 @@ async def full_resource_prompt_context(
         # Collect the list of resources in the results (in order of relevance).
         ordered_resources = []
         for paragraph in ordered_paragraphs:
-            resource_uuid = paragraph.id.split("/")[0]
+            resource_uuid = parse_text_block_id(paragraph.id).rid
             if resource_uuid not in ordered_resources:
-                ordered_resources.append(resource_uuid)
+                skip = False
+                if strategy.apply_to is not None:
+                    # decide whether the resource should be extended or not
+                    for label in strategy.apply_to.exclude:
+                        skip = skip or (label in (paragraph.labels or []))
+
+                if not skip:
+                    ordered_resources.append(resource_uuid)
 
     # For each resource, collect the extracted text from all its fields.
-    resource_extracted_texts = await run_concurrently(
+    resources_extracted_texts = await run_concurrently(
         [
-            get_resource_extracted_texts(kbid, resource_uuid)
-            for resource_uuid in ordered_resources[:number_of_full_resources]
+            hydrate_resource_text(kbid, resource_uuid, max_concurrent_tasks=MAX_RESOURCE_FIELD_TASKS)
+            for resource_uuid in ordered_resources[: strategy.count]
         ],
         max_concurrent=MAX_RESOURCE_TASKS,
     )
-
-    for extracted_texts in resource_extracted_texts:
-        if extracted_texts is None:
+    added_fields = set()
+    for resource_extracted_texts in resources_extracted_texts:
+        if resource_extracted_texts is None:
             continue
-        for field, extracted_text in extracted_texts:
+        for field, extracted_text in resource_extracted_texts:
+            # First off, remove the text block ids from paragraphs that belong to
+            # the same field, as otherwise the context will be duplicated.
+            for tb_id in context.text_block_ids():
+                if tb_id.startswith(field.full()):
+                    del context[tb_id]
             # Add the extracted text of each field to the context.
-            context[field.resource_unique_id] = extracted_text
+            context[field.full()] = extracted_text
+            added_fields.add(field.full())
+
+    if strategy.include_remaining_text_blocks:
+        for paragraph in ordered_paragraphs:
+            pid = cast(ParagraphId, parse_text_block_id(paragraph.id))
+            if pid.field_id.full() not in added_fields:
+                context[paragraph.id] = _clean_paragraph_text(paragraph)
 
 
-async def composed_prompt_context(
+async def extend_prompt_context_with_metadata(
+    context: CappedPromptContext,
+    kbid: str,
+    strategy: MetadataExtensionStrategy,
+) -> None:
+    text_block_ids: list[TextBlockId] = []
+    for text_block_id in context.text_block_ids():
+        try:
+            text_block_ids.append(parse_text_block_id(text_block_id))
+        except ValueError:  # pragma: no cover
+            # Some text block ids are not paragraphs nor fields, so they are skipped
+            # (e.g. USER_CONTEXT_0, when the user provides extra context)
+            continue
+    if len(text_block_ids) == 0:  # pragma: no cover
+        return
+
+    if MetadataExtensionType.ORIGIN in strategy.types:
+        await extend_prompt_context_with_origin_metadata(context, kbid, text_block_ids)
+
+    if MetadataExtensionType.CLASSIFICATION_LABELS in strategy.types:
+        await extend_prompt_context_with_classification_labels(context, kbid, text_block_ids)
+
+    if MetadataExtensionType.NERS in strategy.types:
+        await extend_prompt_context_with_ner(context, kbid, text_block_ids)
+
+    if MetadataExtensionType.EXTRA_METADATA in strategy.types:
+        await extend_prompt_context_with_extra_metadata(context, kbid, text_block_ids)
+
+
+def parse_text_block_id(text_block_id: str) -> TextBlockId:
+    try:
+        # Typically, the text block id is a paragraph id
+        return ParagraphId.from_string(text_block_id)
+    except ValueError:
+        # When we're doing `full_resource` or `hierarchy` strategies,the text block id
+        # is a field id
+        return FieldId.from_string(text_block_id)
+
+
+async def extend_prompt_context_with_origin_metadata(context, kbid, text_block_ids: list[TextBlockId]):
+    async def _get_origin(kbid: str, rid: str) -> tuple[str, Optional[Origin]]:
+        origin = None
+        resource = await cache.get_resource(kbid, rid)
+        if resource is not None:
+            pb_origin = await resource.get_origin()
+            if pb_origin is not None:
+                origin = from_proto.origin(pb_origin)
+        return rid, origin
+
+    rids = {tb_id.rid for tb_id in text_block_ids}
+    origins = await run_concurrently([_get_origin(kbid, rid) for rid in rids])
+    rid_to_origin = {rid: origin for rid, origin in origins if origin is not None}
+    for tb_id in text_block_ids:
+        origin = rid_to_origin.get(tb_id.rid)
+        if origin is not None and tb_id.full() in context.output:
+            context[tb_id.full()] += f"\n\nDOCUMENT METADATA AT ORIGIN:\n{to_yaml(origin)}"
+
+
+async def extend_prompt_context_with_classification_labels(
+    context, kbid, text_block_ids: list[TextBlockId]
+):
+    async def _get_labels(kbid: str, _id: TextBlockId) -> tuple[TextBlockId, list[tuple[str, str]]]:
+        fid = _id if isinstance(_id, FieldId) else _id.field_id
+        labels = set()
+        resource = await cache.get_resource(kbid, fid.rid)
+        if resource is not None:
+            pb_basic = await resource.get_basic()
+            if pb_basic is not None:
+                # Add the classification labels of the resource
+                for classif in pb_basic.usermetadata.classifications:
+                    labels.add((classif.labelset, classif.label))
+                # Add the classifications labels of the field
+                for fc in pb_basic.computedmetadata.field_classifications:
+                    if fc.field.field == fid.key and fc.field.field_type == fid.pb_type:
+                        for classif in fc.classifications:
+                            if classif.cancelled_by_user:  # pragma: no cover
+                                continue
+                            labels.add((classif.labelset, classif.label))
+        return _id, list(labels)
+
+    classif_labels = await run_concurrently([_get_labels(kbid, tb_id) for tb_id in text_block_ids])
+    tb_id_to_labels = {tb_id: labels for tb_id, labels in classif_labels if len(labels) > 0}
+    for tb_id in text_block_ids:
+        labels = tb_id_to_labels.get(tb_id)
+        if labels is not None and tb_id.full() in context.output:
+            labels_text = "DOCUMENT CLASSIFICATION LABELS:"
+            for labelset, label in labels:
+                labels_text += f"\n - {label} ({labelset})"
+            context[tb_id.full()] += "\n\n" + labels_text
+
+
+async def extend_prompt_context_with_ner(context, kbid, text_block_ids: list[TextBlockId]):
+    async def _get_ners(kbid: str, _id: TextBlockId) -> tuple[TextBlockId, dict[str, set[str]]]:
+        fid = _id if isinstance(_id, FieldId) else _id.field_id
+        ners: dict[str, set[str]] = {}
+        resource = await cache.get_resource(kbid, fid.rid)
+        if resource is not None:
+            field = await resource.get_field(fid.key, fid.pb_type, load=False)
+            fcm = await field.get_field_metadata()
+            if fcm is not None:
+                # Data Augmentation + Processor entities
+                for (
+                    data_aumgentation_task_id,
+                    entities_wrapper,
+                ) in fcm.metadata.entities.items():
+                    for entity in entities_wrapper.entities:
+                        ners.setdefault(entity.label, set()).add(entity.text)
+                # Legacy processor entities
+                # TODO: Remove once processor doesn't use this anymore and remove the positions and ner fields from the message
+                for token, family in fcm.metadata.ner.items():
+                    ners.setdefault(family, set()).add(token)
+        return _id, ners
+
+    nerss = await run_concurrently([_get_ners(kbid, tb_id) for tb_id in text_block_ids])
+    tb_id_to_ners = {tb_id: ners for tb_id, ners in nerss if len(ners) > 0}
+    for tb_id in text_block_ids:
+        ners = tb_id_to_ners.get(tb_id)
+        if ners is not None and tb_id.full() in context.output:
+            ners_text = "DOCUMENT NAMED ENTITIES (NERs):"
+            for family, tokens in ners.items():
+                ners_text += f"\n - {family}:"
+                for token in sorted(list(tokens)):
+                    ners_text += f"\n   - {token}"
+            context[tb_id.full()] += "\n\n" + ners_text
+
+
+async def extend_prompt_context_with_extra_metadata(context, kbid, text_block_ids: list[TextBlockId]):
+    async def _get_extra(kbid: str, rid: str) -> tuple[str, Optional[Extra]]:
+        extra = None
+        resource = await cache.get_resource(kbid, rid)
+        if resource is not None:
+            pb_extra = await resource.get_extra()
+            if pb_extra is not None:
+                extra = from_proto.extra(pb_extra)
+        return rid, extra
+
+    rids = {tb_id.rid for tb_id in text_block_ids}
+    extras = await run_concurrently([_get_extra(kbid, rid) for rid in rids])
+    rid_to_extra = {rid: extra for rid, extra in extras if extra is not None}
+    for tb_id in text_block_ids:
+        extra = rid_to_extra.get(tb_id.rid)
+        if extra is not None and tb_id.full() in context.output:
+            context[tb_id.full()] += f"\n\nDOCUMENT EXTRA METADATA:\n{to_yaml(extra)}"
+
+
+def to_yaml(obj: BaseModel) -> str:
+    return yaml.dump(
+        obj.model_dump(exclude_none=True, exclude_defaults=True, exclude_unset=True),
+        default_flow_style=False,
+        indent=2,
+        sort_keys=True,
+    )
+
+
+async def field_extension_prompt_context(
     context: CappedPromptContext,
     kbid: str,
     ordered_paragraphs: list[FindParagraph],
-    extend_with_fields: list[str],
+    strategy: FieldExtensionStrategy,
 ) -> None:
     """
     Algorithm steps:
@@ -310,33 +470,402 @@ async def composed_prompt_context(
     """
     ordered_resources = []
     for paragraph in ordered_paragraphs:
-        resource_uuid = paragraph.id.split("/")[0]
+        resource_uuid = ParagraphId.from_string(paragraph.id).rid
         if resource_uuid not in ordered_resources:
             ordered_resources.append(resource_uuid)
 
     # Fetch the extracted texts of the specified fields for each resource
-    txn = await get_read_only_transaction()
-    kb_obj = KnowledgeBoxORM(txn, await get_storage(), kbid)
-
-    tasks = [
-        get_resource_field_extracted_text(kb_obj, resource_uuid, field_id)
-        for resource_uuid in ordered_resources
-        for field_id in extend_with_fields
-    ]
+    extend_fields = strategy.fields
+    extend_field_ids = []
+    for resource_uuid in ordered_resources:
+        for field_id in extend_fields:
+            try:
+                fid = FieldId.from_string(f"{resource_uuid}/{field_id.strip('/')}")
+                extend_field_ids.append(fid)
+            except ValueError:  # pragma: no cover
+                # Invalid field id, skiping
+                continue
+
+    tasks = [hydrate_field_text(kbid, fid) for fid in extend_field_ids]
     field_extracted_texts = await run_concurrently(tasks)
 
     for result in field_extracted_texts:
-        if result is None:
+        if result is None:  # pragma: no cover
             continue
-        # Add the extracted text of each field to the beginning of the context.
         field, extracted_text = result
-        context[field.resource_unique_id] = extracted_text
+        # First off, remove the text block ids from paragraphs that belong to
+        # the same field, as otherwise the context will be duplicated.
+        for tb_id in context.text_block_ids():
+            if tb_id.startswith(field.full()):
+                del context[tb_id]
+        # Add the extracted text of each field to the beginning of the context.
+        context[field.full()] = extracted_text
 
     # Add the extracted text of each paragraph to the end of the context.
     for paragraph in ordered_paragraphs:
         context[paragraph.id] = _clean_paragraph_text(paragraph)
 
 
+async def get_paragraph_text_with_neighbours(
+    kbid: str,
+    pid: ParagraphId,
+    field_paragraphs: list[ParagraphId],
+    before: int = 0,
+    after: int = 0,
+) -> tuple[ParagraphId, str]:
+    """
+    This function will get the paragraph text of the paragraph with the neighbouring paragraphs included.
+    Parameters:
+        kbid: The knowledge box id.
+        pid: The matching paragraph id.
+        field_paragraphs: The list of paragraph ids of the field.
+        before: The number of paragraphs to include before the matching paragraph.
+        after: The number of paragraphs to include after the matching paragraph.
+    """
+
+    async def _get_paragraph_text(
+        kbid: str,
+        pid: ParagraphId,
+    ) -> tuple[ParagraphId, str]:
+        return pid, await get_paragraph_text(
+            kbid=kbid,
+            paragraph_id=pid,
+            log_on_missing_field=True,
+        )
+
+    ops = []
+    try:
+        for paragraph_index in get_neighbouring_paragraph_indexes(
+            field_paragraphs=field_paragraphs,
+            matching_paragraph=pid,
+            before=before,
+            after=after,
+        ):
+            neighbour_pid = field_paragraphs[paragraph_index]
+            ops.append(
+                asyncio.create_task(
+                    _get_paragraph_text(
+                        kbid=kbid,
+                        pid=neighbour_pid,
+                    )
+                )
+            )
+    except ParagraphIdNotFoundInExtractedMetadata:
+        logger.warning(
+            "Could not find matching paragraph in extracted metadata. This is odd and needs to be investigated.",
+            extra={
+                "kbid": kbid,
+                "matching_paragraph": pid.full(),
+                "field_paragraphs": [p.full() for p in field_paragraphs],
+            },
+        )
+        # If we could not find the matching paragraph in the extracted metadata, we can't retrieve
+        # the neighbouring paragraphs and we simply fetch the text of the matching paragraph.
+        ops.append(
+            asyncio.create_task(
+                _get_paragraph_text(
+                    kbid=kbid,
+                    pid=pid,
+                )
+            )
+        )
+
+    results = []
+    if len(ops) > 0:
+        results = await asyncio.gather(*ops)
+
+    # Sort the results by the paragraph start
+    results.sort(key=lambda x: x[0].paragraph_start)
+    paragraph_texts = []
+    for _, text in results:
+        if text != "":
+            paragraph_texts.append(text)
+    return pid, "\n\n".join(paragraph_texts)
+
+
+async def get_field_paragraphs_list(
+    kbid: str,
+    field: FieldId,
+    paragraphs: list[ParagraphId],
+) -> None:
+    """
+    Modifies the paragraphs list by adding the paragraph ids of the field, sorted by position.
+    """
+    resource = await cache.get_resource(kbid, field.rid)
+    if resource is None:  # pragma: no cover
+        return
+    field_obj: Field = await resource.get_field(key=field.key, type=field.pb_type, load=False)
+    field_metadata: Optional[resources_pb2.FieldComputedMetadata] = await field_obj.get_field_metadata(
+        force=True
+    )
+    if field_metadata is None:  # pragma: no cover
+        return
+    for paragraph in field_metadata.metadata.paragraphs:
+        paragraphs.append(
+            ParagraphId(
+                field_id=field,
+                paragraph_start=paragraph.start,
+                paragraph_end=paragraph.end,
+            )
+        )
+
+
+async def neighbouring_paragraphs_prompt_context(
+    context: CappedPromptContext,
+    kbid: str,
+    ordered_text_blocks: list[FindParagraph],
+    strategy: NeighbouringParagraphsStrategy,
+) -> None:
+    """
+    This function will get the paragraph texts and then craft a context with the neighbouring paragraphs of the
+    paragraphs in the ordered_paragraphs list. The number of paragraphs to include before and after each paragraph
+    """
+    # First, get the sorted list of paragraphs for each matching field
+    # so we can know the indexes of the neighbouring paragraphs
+    unique_fields = {
+        ParagraphId.from_string(text_block.id).field_id for text_block in ordered_text_blocks
+    }
+    paragraphs_by_field: dict[FieldId, list[ParagraphId]] = {}
+    field_ops = []
+    for field_id in unique_fields:
+        plist = paragraphs_by_field.setdefault(field_id, [])
+        field_ops.append(
+            asyncio.create_task(get_field_paragraphs_list(kbid=kbid, field=field_id, paragraphs=plist))
+        )
+    if field_ops:
+        await asyncio.gather(*field_ops)
+
+    # Now, get the paragraph texts with the neighbouring paragraphs
+    paragraph_ops = []
+    for text_block in ordered_text_blocks:
+        pid = ParagraphId.from_string(text_block.id)
+        paragraph_ops.append(
+            asyncio.create_task(
+                get_paragraph_text_with_neighbours(
+                    kbid=kbid,
+                    pid=pid,
+                    before=strategy.before,
+                    after=strategy.after,
+                    field_paragraphs=paragraphs_by_field.get(pid.field_id, []),
+                )
+            )
+        )
+    if not paragraph_ops:  # pragma: no cover
+        return
+
+    results: list[tuple[ParagraphId, str]] = await asyncio.gather(*paragraph_ops)
+    # Add the paragraph texts to the context
+    for pid, text in results:
+        if text != "":
+            context[pid.full()] = text
+
+
+async def conversation_prompt_context(
+    context: CappedPromptContext,
+    kbid: str,
+    ordered_paragraphs: list[FindParagraph],
+    conversational_strategy: ConversationalStrategy,
+    visual_llm: bool,
+):
+    analyzed_fields: List[str] = []
+    async with get_driver().transaction(read_only=True) as txn:
+        storage = await get_storage()
+        kb = KnowledgeBoxORM(txn, storage, kbid)
+        for paragraph in ordered_paragraphs:
+            context[paragraph.id] = _clean_paragraph_text(paragraph)
+
+            # If the paragraph is a conversation and it matches semantically, we assume we
+            # have matched with the question, therefore try to include the answer to the
+            # context by pulling the next few messages of the conversation field
+            rid, field_type, field_id, mident = paragraph.id.split("/")[:4]
+            if field_type == "c" and paragraph.score_type in (
+                SCORE_TYPE.VECTOR,
+                SCORE_TYPE.BOTH,
+                SCORE_TYPE.BM25,
+            ):
+                field_unique_id = "-".join([rid, field_type, field_id])
+                if field_unique_id in analyzed_fields:
+                    continue
+                resource = await kb.get(rid)
+                if resource is None:  # pragma: no cover
+                    continue
+
+                field_obj: Conversation = await resource.get_field(
+                    field_id, FIELD_TYPE_STR_TO_PB["c"], load=True
+                )  # type: ignore
+                cmetadata = await field_obj.get_metadata()
+
+                attachments: List[resources_pb2.FieldRef] = []
+                if conversational_strategy.full:
+                    extracted_text = await field_obj.get_extracted_text()
+                    for current_page in range(1, cmetadata.pages + 1):
+                        conv = await field_obj.db_get_value(current_page)
+
+                        for message in conv.messages:
+                            ident = message.ident
+                            if extracted_text is not None:
+                                text = extracted_text.split_text.get(ident, message.content.text.strip())
+                            else:
+                                text = message.content.text.strip()
+                            pid = f"{rid}/{field_type}/{field_id}/{ident}/0-{len(text) + 1}"
+                            context[pid] = text
+                            attachments.extend(message.content.attachments_fields)
+                else:
+                    # Add first message
+                    extracted_text = await field_obj.get_extracted_text()
+                    first_page = await field_obj.db_get_value()
+                    if len(first_page.messages) > 0:
+                        message = first_page.messages[0]
+                        ident = message.ident
+                        if extracted_text is not None:
+                            text = extracted_text.split_text.get(ident, message.content.text.strip())
+                        else:
+                            text = message.content.text.strip()
+                        pid = f"{rid}/{field_type}/{field_id}/{ident}/0-{len(text) + 1}"
+                        context[pid] = text
+                        attachments.extend(message.content.attachments_fields)
+
+                    messages: Deque[resources_pb2.Message] = deque(
+                        maxlen=conversational_strategy.max_messages
+                    )
+
+                    pending = -1
+                    for page in range(1, cmetadata.pages + 1):
+                        # Collect the messages with the window asked by the user arround the match paragraph
+                        conv = await field_obj.db_get_value(page)
+                        for message in conv.messages:
+                            messages.append(message)
+                            if pending > 0:
+                                pending -= 1
+                            if message.ident == mident:
+                                pending = (conversational_strategy.max_messages - 1) // 2
+                            if pending == 0:
+                                break
+                        if pending == 0:
+                            break
+
+                    for message in messages:
+                        text = message.content.text.strip()
+                        pid = f"{rid}/{field_type}/{field_id}/{message.ident}/0-{len(message.content.text) + 1}"
+                        context[pid] = text
+                        attachments.extend(message.content.attachments_fields)
+
+                if conversational_strategy.attachments_text:
+                    # add on the context the images if vlm enabled
+                    for attachment in attachments:
+                        field: File = await resource.get_field(
+                            attachment.field_id, attachment.field_type, load=True
+                        )  # type: ignore
+                        extracted_text = await field.get_extracted_text()
+                        if extracted_text is not None:
+                            pid = f"{rid}/{field_type}/{attachment.field_id}/0-{len(extracted_text.text) + 1}"
+                            context[pid] = f"Attachment {attachment.field_id}: {extracted_text.text}\n\n"
+
+                if conversational_strategy.attachments_images and visual_llm:
+                    for attachment in attachments:
+                        file_field: File = await resource.get_field(
+                            attachment.field_id, attachment.field_type, load=True
+                        )  # type: ignore
+                        image = await get_file_thumbnail_image(file_field)
+                        if image is not None:
+                            pid = f"{rid}/f/{attachment.field_id}/0-0"
+                            context.images[pid] = image
+
+                analyzed_fields.append(field_unique_id)
+
+
+async def hierarchy_prompt_context(
+    context: CappedPromptContext,
+    kbid: str,
+    ordered_paragraphs: list[FindParagraph],
+    strategy: HierarchyResourceStrategy,
+) -> None:
+    """
+    This function will get the paragraph texts (possibly with extra characters, if extra_characters > 0) and then
+    craft a context with all paragraphs of the same resource grouped together. Moreover, on each group of paragraphs,
+    it includes the resource title and summary so that the LLM can have a better understanding of the context.
+    """
+    paragraphs_extra_characters = max(strategy.count, 0)
+    # Make a copy of the ordered paragraphs to avoid modifying the original list, which is returned
+    # in the response to the user
+    ordered_paragraphs_copy = copy.deepcopy(ordered_paragraphs)
+    resources: Dict[str, ExtraCharsParagraph] = {}
+
+    # Iterate paragraphs to get extended text
+    for paragraph in ordered_paragraphs_copy:
+        paragraph_id = ParagraphId.from_string(paragraph.id)
+        extended_paragraph_text = paragraph.text
+        if paragraphs_extra_characters > 0:
+            extended_paragraph_text = await get_paragraph_text(
+                kbid=kbid,
+                paragraph_id=paragraph_id,
+                log_on_missing_field=True,
+            )
+        rid = paragraph_id.rid
+        if rid not in resources:
+            # Get the title and the summary of the resource
+            title_text = await get_paragraph_text(
+                kbid=kbid,
+                paragraph_id=ParagraphId(
+                    field_id=FieldId(
+                        rid=rid,
+                        type="a",
+                        key="title",
+                    ),
+                    paragraph_start=0,
+                    paragraph_end=500,
+                ),
+                log_on_missing_field=False,
+            )
+            summary_text = await get_paragraph_text(
+                kbid=kbid,
+                paragraph_id=ParagraphId(
+                    field_id=FieldId(
+                        rid=rid,
+                        type="a",
+                        key="summary",
+                    ),
+                    paragraph_start=0,
+                    paragraph_end=1000,
+                ),
+                log_on_missing_field=False,
+            )
+            resources[rid] = ExtraCharsParagraph(
+                title=title_text,
+                summary=summary_text,
+                paragraphs=[(paragraph, extended_paragraph_text)],
+            )
+        else:
+            resources[rid].paragraphs.append((paragraph, extended_paragraph_text))
+
+    # Modify the first paragraph of each resource to include the title and summary of the resource, as well as the
+    # extended paragraph text of all the paragraphs in the resource.
+    for values in resources.values():
+        title_text = values.title
+        summary_text = values.summary
+        first_paragraph = None
+        text_with_hierarchy = ""
+        for paragraph, extended_paragraph_text in values.paragraphs:
+            if first_paragraph is None:
+                first_paragraph = paragraph
+            text_with_hierarchy += "\n EXTRACTED BLOCK: \n " + extended_paragraph_text + " \n\n "
+            # All paragraphs of the resource are cleared except the first one, which will be the
+            # one containing the whole hierarchy information
+            paragraph.text = ""
+
+        if first_paragraph is not None:
+            # The first paragraph is the only one holding the hierarchy information
+            first_paragraph.text = f"DOCUMENT: {title_text} \n SUMMARY: {summary_text} \n RESOURCE CONTENT: {text_with_hierarchy}"
+
+    # Now that the paragraphs have been modified, we can add them to the context
+    for paragraph in ordered_paragraphs_copy:
+        if paragraph.text == "":
+            # Skip paragraphs that were cleared in the hierarchy expansion
+            continue
+        context[paragraph.id] = _clean_paragraph_text(paragraph)
+    return
+
+
 class PromptContextBuilder:
     """
     Builds the context for the LLM prompt.
@@ -345,7 +874,7 @@ class PromptContextBuilder:
     def __init__(
         self,
         kbid: str,
-        find_results: KnowledgeboxFindResults,
+        ordered_paragraphs: list[FindParagraph],
         resource: Optional[str] = None,
         user_context: Optional[list[str]] = None,
         strategies: Optional[Sequence[RagStrategy]] = None,
@@ -354,7 +883,7 @@ class PromptContextBuilder:
         visual_llm: bool = False,
     ):
         self.kbid = kbid
-        self.ordered_paragraphs = get_ordered_paragraphs(find_results)
+        self.ordered_paragraphs = ordered_paragraphs
         self.resource = resource
         self.user_context = user_context
         self.strategies = strategies
@@ -374,98 +903,175 @@ class PromptContextBuilder:
         ccontext = CappedPromptContext(max_size=self.max_context_characters)
         self.prepend_user_context(ccontext)
         await self._build_context(ccontext)
-
         if self.visual_llm:
             await self._build_context_images(ccontext)
 
         context = ccontext.output
         context_images = ccontext.images
-        context_order = {
-            text_block_id: order for order, text_block_id in enumerate(context.keys())
-        }
+        context_order = {text_block_id: order for order, text_block_id in enumerate(context.keys())}
         return context, context_order, context_images
 
     async def _build_context_images(self, context: CappedPromptContext) -> None:
-        flatten_strategies = []
-        page_count = 5
-        gather_pages = False
-        gather_tables = False
-        if self.image_strategies is not None:
-            for strategy in self.image_strategies:
-                flatten_strategies.append(strategy.name)
-                if strategy.name == ImageRagStrategyName.PAGE_IMAGE:
-                    gather_pages = True
-                    if strategy.count is not None:  # type: ignore
-                        page_count = strategy.count  # type: ignore
-                if strategy.name == ImageRagStrategyName.TABLES:
-                    gather_tables = True
-
+        if self.image_strategies is None or len(self.image_strategies) == 0:
+            # Nothing to do
+            return
+        page_image_strategy: Optional[PageImageStrategy] = None
+        max_page_images = 5
+        table_image_strategy: Optional[TableImageStrategy] = None
+        paragraph_image_strategy: Optional[ParagraphImageStrategy] = None
+        for strategy in self.image_strategies:
+            if strategy.name == ImageRagStrategyName.PAGE_IMAGE:
+                if page_image_strategy is None:
+                    page_image_strategy = cast(PageImageStrategy, strategy)
+                    if page_image_strategy.count is not None:
+                        max_page_images = page_image_strategy.count
+            elif strategy.name == ImageRagStrategyName.TABLES:
+                if table_image_strategy is None:
+                    table_image_strategy = cast(TableImageStrategy, strategy)
+            elif strategy.name == ImageRagStrategyName.PARAGRAPH_IMAGE:
+                if paragraph_image_strategy is None:
+                    paragraph_image_strategy = cast(ParagraphImageStrategy, strategy)
+            else:  # pragma: no cover
+                logger.warning(
+                    "Unknown image strategy",
+                    extra={"strategy": strategy.name, "kbid": self.kbid},
+                )
+        page_images_added = 0
         for paragraph in self.ordered_paragraphs:
-            if paragraph.page_with_visual and paragraph.position:
-                if (
-                    gather_pages
-                    and paragraph.position.page_number
-                    and len(context.images) < page_count
-                ):
-                    field = "/".join(paragraph.id.split("/")[:3])
-                    page = paragraph.position.page_number
-                    page_id = f"{field}/{page}"
-                    if page_id not in context.images:
-                        context.images[page_id] = await get_page_image(
-                            self.kbid, paragraph.id, page
-                        )
+            pid = ParagraphId.from_string(paragraph.id)
+            paragraph_page_number = get_paragraph_page_number(paragraph)
             if (
-                gather_tables
-                and paragraph.is_a_table
-                and paragraph.reference
-                and paragraph.reference != ""
+                page_image_strategy is not None
+                and page_images_added < max_page_images
+                and paragraph_page_number is not None
             ):
-                image = paragraph.reference
-                context.images[paragraph.id] = await get_paragraph_image(
-                    self.kbid, paragraph.id, image
-                )
+                # page_image_id: rid/f/myfield/0
+                page_image_id = "/".join([pid.field_id.full(), str(paragraph_page_number)])
+                if page_image_id not in context.images:
+                    image = await get_page_image(self.kbid, pid, paragraph_page_number)
+                    if image is not None:
+                        context.images[page_image_id] = image
+                        page_images_added += 1
+                    else:
+                        logger.warning(
+                            f"Could not retrieve image for paragraph from storage",
+                            extra={
+                                "kbid": self.kbid,
+                                "paragraph": pid.full(),
+                                "page_number": paragraph_page_number,
+                            },
+                        )
+
+            add_table = table_image_strategy is not None and paragraph.is_a_table
+            add_paragraph = paragraph_image_strategy is not None and not paragraph.is_a_table
+            if (add_table or add_paragraph) and (
+                paragraph.reference is not None and paragraph.reference != ""
+            ):
+                pimage = await get_paragraph_image(self.kbid, pid, paragraph.reference)
+                if pimage is not None:
+                    context.images[paragraph.id] = pimage
+                else:
+                    logger.warning(
+                        f"Could not retrieve image for paragraph from storage",
+                        extra={
+                            "kbid": self.kbid,
+                            "paragraph": pid.full(),
+                            "reference": paragraph.reference,
+                        },
+                    )
 
     async def _build_context(self, context: CappedPromptContext) -> None:
         if self.strategies is None or len(self.strategies) == 0:
+            # When no strategy is specified, use the default one
             await default_prompt_context(context, self.kbid, self.ordered_paragraphs)
             return
-
-        number_of_full_resources = 0
-        distance = 0
-        extend_with_fields = []
+        else:
+            # Add the paragraphs to the context and then apply the strategies
+            for paragraph in self.ordered_paragraphs:
+                context[paragraph.id] = _clean_paragraph_text(paragraph)
+
+        full_resource: Optional[FullResourceStrategy] = None
+        hierarchy: Optional[HierarchyResourceStrategy] = None
+        neighbouring_paragraphs: Optional[NeighbouringParagraphsStrategy] = None
+        field_extension: Optional[FieldExtensionStrategy] = None
+        metadata_extension: Optional[MetadataExtensionStrategy] = None
+        conversational_strategy: Optional[ConversationalStrategy] = None
         for strategy in self.strategies:
             if strategy.name == RagStrategyName.FIELD_EXTENSION:
-                extend_with_fields.extend(strategy.fields)  # type: ignore
+                field_extension = cast(FieldExtensionStrategy, strategy)
+            elif strategy.name == RagStrategyName.CONVERSATION:
+                conversational_strategy = cast(ConversationalStrategy, strategy)
             elif strategy.name == RagStrategyName.FULL_RESOURCE:
-                if self.resource:
-                    number_of_full_resources = 1
-                else:
-                    number_of_full_resources = strategy.count or len(self.ordered_paragraphs)  # type: ignore
+                full_resource = cast(FullResourceStrategy, strategy)
+                if self.resource:  # pragma: no cover
+                    # When the retrieval is scoped to a specific resource
+                    # the full resource strategy only includes that resource
+                    full_resource.count = 1
             elif strategy.name == RagStrategyName.HIERARCHY:
-                distance = strategy.count  # type: ignore
+                hierarchy = cast(HierarchyResourceStrategy, strategy)
+            elif strategy.name == RagStrategyName.NEIGHBOURING_PARAGRAPHS:
+                neighbouring_paragraphs = cast(NeighbouringParagraphsStrategy, strategy)
+            elif strategy.name == RagStrategyName.METADATA_EXTENSION:
+                metadata_extension = cast(MetadataExtensionStrategy, strategy)
+            elif strategy.name != RagStrategyName.PREQUERIES:  # pragma: no cover
+                # Prequeries are not handled here
+                logger.warning(
+                    "Unknown rag strategy",
+                    extra={"strategy": strategy.name, "kbid": self.kbid},
+                )
 
-        if number_of_full_resources:
+        if full_resource:
+            # When full resoure is enabled, only metadata extension is allowed.
             await full_resource_prompt_context(
                 context,
                 self.kbid,
                 self.ordered_paragraphs,
                 self.resource,
-                number_of_full_resources,
+                full_resource,
             )
+            if metadata_extension:
+                await extend_prompt_context_with_metadata(context, self.kbid, metadata_extension)
             return
 
-        if distance > 0:
-            await get_extra_chars(self.kbid, self.ordered_paragraphs, distance)
-            await default_prompt_context(context, self.kbid, self.ordered_paragraphs)
-            return
+        if hierarchy:
+            await hierarchy_prompt_context(
+                context,
+                self.kbid,
+                self.ordered_paragraphs,
+                hierarchy,
+            )
+        if neighbouring_paragraphs:
+            await neighbouring_paragraphs_prompt_context(
+                context,
+                self.kbid,
+                self.ordered_paragraphs,
+                neighbouring_paragraphs,
+            )
+        if field_extension:
+            await field_extension_prompt_context(
+                context,
+                self.kbid,
+                self.ordered_paragraphs,
+                field_extension,
+            )
+        if conversational_strategy:
+            await conversation_prompt_context(
+                context,
+                self.kbid,
+                self.ordered_paragraphs,
+                conversational_strategy,
+                self.visual_llm,
+            )
+        if metadata_extension:
+            await extend_prompt_context_with_metadata(context, self.kbid, metadata_extension)
 
-        await composed_prompt_context(
-            context,
-            self.kbid,
-            self.ordered_paragraphs,
-            extend_with_fields=extend_with_fields,
-        )
-        return
+
+def get_paragraph_page_number(paragraph: FindParagraph) -> Optional[int]:
+    if not paragraph.page_with_visual:
+        return None
+    if paragraph.position is None:
+        return None
+    return paragraph.position.page_number
 
 
 @dataclass
@@ -475,67 +1081,6 @@ class ExtraCharsParagraph:
     paragraphs: List[Tuple[FindParagraph, str]]
 
 
-async def get_extra_chars(
-    kbid: str, ordered_paragraphs: list[FindParagraph], distance: int
-):
-    etcache = paragraphs.ExtractedTextCache()
-    resources: Dict[str, ExtraCharsParagraph] = {}
-    for paragraph in ordered_paragraphs:
-        rid, field_type, field = paragraph.id.split("/")[:3]
-        field_path = "/".join([rid, field_type, field])
-        position = paragraph.id.split("/")[-1]
-        start, end = position.split("-")
-        int_start = int(start)
-        int_end = int(end) + distance
-
-        new_text = await paragraphs.get_paragraph_text(
-            kbid=kbid,
-            rid=rid,
-            field=field_path,
-            start=int_start,
-            end=int_end,
-            extracted_text_cache=etcache,
-        )
-        if rid not in resources:
-            title_text = await paragraphs.get_paragraph_text(
-                kbid=kbid,
-                rid=rid,
-                field="/a/title",
-                start=0,
-                end=500,
-                extracted_text_cache=etcache,
-            )
-            summary_text = await paragraphs.get_paragraph_text(
-                kbid=kbid,
-                rid=rid,
-                field="/a/summary",
-                start=0,
-                end=1000,
-                extracted_text_cache=etcache,
-            )
-            resources[rid] = ExtraCharsParagraph(
-                title=title_text,
-                summary=summary_text,
-                paragraphs=[(paragraph, new_text)],
-            )
-        else:
-            resources[rid].paragraphs.append((paragraph, new_text))  # type: ignore
-
-    for values in resources.values():
-        title_text = values.title
-        summary_text = values.summary
-        first_paragraph = None
-        text = ""
-        for paragraph, text in values.paragraphs:
-            if first_paragraph is None:
-                first_paragraph = paragraph
-            text += "EXTRACTED BLOCK: \n " + text + " \n\n "
-            paragraph.text = ""
-
-        if first_paragraph is not None:
-            first_paragraph.text = f"DOCUMENT: {title_text} \n SUMMARY: {summary_text} \n RESOURCE CONTENT: {text}"
-
-
 def _clean_paragraph_text(paragraph: FindParagraph) -> str:
     text = paragraph.text.strip()
     # Do not send highlight marks on prompt context
@@ -543,17 +1088,23 @@ def _clean_paragraph_text(paragraph: FindParagraph) -> str:
     return text
 
 
-def get_ordered_paragraphs(results: KnowledgeboxFindResults) -> list[FindParagraph]:
+def get_neighbouring_paragraph_indexes(
+    field_paragraphs: list[ParagraphId],
+    matching_paragraph: ParagraphId,
+    before: int,
+    after: int,
+) -> list[int]:
     """
-    Returns the list of paragraphs in the results, ordered by relevance.
+    Returns the indexes of the neighbouring paragraphs to fetch (including the matching paragraph).
     """
-    return sorted(
-        [
-            paragraph
-            for resource in results.resources.values()
-            for field in resource.fields.values()
-            for paragraph in field.paragraphs.values()
-        ],
-        key=lambda paragraph: paragraph.order,
-        reverse=False,
-    )
+    assert before >= 0
+    assert after >= 0
+    try:
+        matching_index = field_paragraphs.index(matching_paragraph)
+    except ValueError:
+        raise ParagraphIdNotFoundInExtractedMetadata(
+            f"Matching paragraph {matching_paragraph.full()} not found in extracted metadata"
+        )
+    start_index = max(0, matching_index - before)
+    end_index = min(len(field_paragraphs), matching_index + after + 1)
+    return list(range(start_index, end_index))