from hashlib import md5 from typing import Any, Dict, List, Optional from langchain_core.utils import get_from_dict_or_env from langchain_community.graphs.graph_document import GraphDocument from langchain_community.graphs.graph_store import GraphStore BASE_ENTITY_LABEL = "__Entity__" EXCLUDED_LABELS = ["_Bloom_Perspective_", "_Bloom_Scene_"] EXCLUDED_RELS = ["_Bloom_HAS_SCENE_"] EXHAUSTIVE_SEARCH_LIMIT = 10000 LIST_LIMIT = 128 # Threshold for returning all available prop values in graph schema DISTINCT_VALUE_LIMIT = 10 node_properties_query = """ CALL apoc.meta.data() YIELD label, other, elementType, type, property WHERE NOT type = "RELATIONSHIP" AND elementType = "node" AND NOT label IN $EXCLUDED_LABELS WITH label AS nodeLabels, collect({property:property, type:type}) AS properties RETURN {labels: nodeLabels, properties: properties} AS output """ rel_properties_query = """ CALL apoc.meta.data() YIELD label, other, elementType, type, property WHERE NOT type = "RELATIONSHIP" AND elementType = "relationship" AND NOT label in $EXCLUDED_LABELS WITH label AS nodeLabels, collect({property:property, type:type}) AS properties RETURN {type: nodeLabels, properties: properties} AS output """ rel_query = """ CALL apoc.meta.data() YIELD label, other, elementType, type, property WHERE type = "RELATIONSHIP" AND elementType = "node" UNWIND other AS other_node WITH * WHERE NOT label IN $EXCLUDED_LABELS AND NOT other_node IN $EXCLUDED_LABELS RETURN {start: label, type: property, end: toString(other_node)} AS output """ include_docs_query = ( "MERGE (d:Document {id:$document.metadata.id}) " "SET d.text = $document.page_content " "SET d += $document.metadata " "WITH d " ) def clean_string_values(text: str) -> str: return text.replace("\n", " ").replace("\r", " ") def value_sanitize(d: Any) -> Any: """Sanitize the input dictionary or list. Sanitizes the input by removing embedding-like values, lists with more than 128 elements, that are mostly irrelevant for generating answers in a LLM context. These properties, if left in results, can occupy significant context space and detract from the LLM's performance by introducing unnecessary noise and cost. """ if isinstance(d, dict): new_dict = {} for key, value in d.items(): if isinstance(value, dict): sanitized_value = value_sanitize(value) if ( sanitized_value is not None ): # Check if the sanitized value is not None new_dict[key] = sanitized_value elif isinstance(value, list): if len(value) < LIST_LIMIT: sanitized_value = value_sanitize(value) if ( sanitized_value is not None ): # Check if the sanitized value is not None new_dict[key] = sanitized_value # Do not include the key if the list is oversized else: new_dict[key] = value return new_dict elif isinstance(d, list): if len(d) < LIST_LIMIT: return [ value_sanitize(item) for item in d if value_sanitize(item) is not None ] else: return None else: return d def _get_node_import_query(baseEntityLabel: bool, include_source: bool) -> str: if baseEntityLabel: return ( f"{include_docs_query if include_source else ''}" "UNWIND $data AS row " f"MERGE (source:`{BASE_ENTITY_LABEL}` {{id: row.id}}) " "SET source += row.properties " f"{'MERGE (d)-[:MENTIONS]->(source) ' if include_source else ''}" "WITH source, row " "CALL apoc.create.addLabels( source, [row.type] ) YIELD node " "RETURN distinct 'done' AS result" ) else: return ( f"{include_docs_query if include_source else ''}" "UNWIND $data AS row " "CALL apoc.merge.node([row.type], {id: row.id}, " "row.properties, {}) YIELD node " f"{'MERGE (d)-[:MENTIONS]->(node) ' if include_source else ''}" "RETURN distinct 'done' AS result" ) def _get_rel_import_query(baseEntityLabel: bool) -> str: if baseEntityLabel: return ( "UNWIND $data AS row " f"MERGE (source:`{BASE_ENTITY_LABEL}` {{id: row.source}}) " f"MERGE (target:`{BASE_ENTITY_LABEL}` {{id: row.target}}) " "WITH source, target, row " "CALL apoc.merge.relationship(source, row.type, " "{}, row.properties, target) YIELD rel " "RETURN distinct 'done'" ) else: return ( "UNWIND $data AS row " "CALL apoc.merge.node([row.source_label], {id: row.source}," "{}, {}) YIELD node as source " "CALL apoc.merge.node([row.target_label], {id: row.target}," "{}, {}) YIELD node as target " "CALL apoc.merge.relationship(source, row.type, " "{}, row.properties, target) YIELD rel " "RETURN distinct 'done'" ) def _enhanced_schema_cypher( label_or_type: str, properties: List[Dict[str, Any]], exhaustive: bool, is_relationship: bool = False, ) -> str: if is_relationship: match_clause = f"MATCH ()-[n:{label_or_type}]->()" else: match_clause = f"MATCH (n:{label_or_type})" with_clauses = [] return_clauses = [] output_dict = {} if exhaustive: for prop in properties: prop_name = prop["property"] prop_type = prop["type"] if prop_type == "STRING": with_clauses.append( ( f"collect(distinct substring(n.`{prop_name}`, 0, 50)) " f"AS `{prop_name}_values`" ) ) return_clauses.append( ( f"values:`{prop_name}_values`[..{DISTINCT_VALUE_LIMIT}]," f" distinct_count: size(`{prop_name}_values`)" ) ) elif prop_type in [ "INTEGER", "FLOAT", "DATE", "DATE_TIME", "LOCAL_DATE_TIME", ]: with_clauses.append(f"min(n.`{prop_name}`) AS `{prop_name}_min`") with_clauses.append(f"max(n.`{prop_name}`) AS `{prop_name}_max`") with_clauses.append( f"count(distinct n.`{prop_name}`) AS `{prop_name}_distinct`" ) return_clauses.append( ( f"min: toString(`{prop_name}_min`), " f"max: toString(`{prop_name}_max`), " f"distinct_count: `{prop_name}_distinct`" ) ) elif prop_type == "LIST": with_clauses.append( ( f"min(size(n.`{prop_name}`)) AS `{prop_name}_size_min`, " f"max(size(n.`{prop_name}`)) AS `{prop_name}_size_max`" ) ) return_clauses.append( f"min_size: `{prop_name}_size_min`, " f"max_size: `{prop_name}_size_max`" ) elif prop_type in ["BOOLEAN", "POINT", "DURATION"]: continue output_dict[prop_name] = "{" + return_clauses.pop() + "}" else: # Just sample 5 random nodes match_clause += " WITH n LIMIT 5" for prop in properties: prop_name = prop["property"] prop_type = prop["type"] if prop_type == "STRING": with_clauses.append( ( f"collect(distinct substring(n.`{prop_name}`, 0, 50)) " f"AS `{prop_name}_values`" ) ) return_clauses.append(f"values: `{prop_name}_values`") elif prop_type in [ "INTEGER", "FLOAT", "DATE", "DATE_TIME", "LOCAL_DATE_TIME", ]: with_clauses.append( f"collect(distinct toString(n.`{prop_name}`)) " f"AS `{prop_name}_values`" ) return_clauses.append(f"values: `{prop_name}_values`") elif prop_type == "LIST": with_clauses.append( ( f"min(size(n.`{prop_name}`)) AS `{prop_name}_size_min`, " f"max(size(n.`{prop_name}`)) AS `{prop_name}_size_max`" ) ) return_clauses.append( f"min_size: `{prop_name}_size_min`,max_size: `{prop_name}_size_max`" ) elif prop_type in ["BOOLEAN", "POINT", "DURATION"]: continue output_dict[prop_name] = "{" + return_clauses.pop() + "}" with_clause = "WITH " + ",\n ".join(with_clauses) return_clause = ( "RETURN {" + ", ".join(f"`{k}`: {v}" for k, v in output_dict.items()) + "} AS output" ) # Combine all parts of the Cypher query cypher_query = "\n".join([match_clause, with_clause, return_clause]) return cypher_query def _format_schema(schema: Dict, is_enhanced: bool) -> str: formatted_node_props = [] formatted_rel_props = [] if is_enhanced: # Enhanced formatting for nodes for node_type, properties in schema["node_props"].items(): formatted_node_props.append(f"- **{node_type}**") for prop in properties: example = "" if prop["type"] == "STRING": if prop.get("distinct_count", 11) > DISTINCT_VALUE_LIMIT: example = ( f'Example: "{clean_string_values(prop["values"][0])}"' if prop["values"] else "" ) else: # If less than 10 possible values return all example = ( ( "Available options: " f'{[clean_string_values(el) for el in prop["values"]]}' ) if prop["values"] else "" ) elif prop["type"] in [ "INTEGER", "FLOAT", "DATE", "DATE_TIME", "LOCAL_DATE_TIME", ]: if prop.get("min") is not None: example = f'Min: {prop["min"]}, Max: {prop["max"]}' else: example = ( f'Example: "{prop["values"][0]}"' if prop["values"] else "" ) elif prop["type"] == "LIST": # Skip embeddings if prop["min_size"] > LIST_LIMIT: continue example = ( f'Min Size: {prop["min_size"]}, Max Size: {prop["max_size"]}' ) formatted_node_props.append( f" - `{prop['property']}`: {prop['type']}` {example}" ) # Enhanced formatting for relationships for rel_type, properties in schema["rel_props"].items(): formatted_rel_props.append(f"- **{rel_type}**") for prop in properties: example = "" if prop["type"] == "STRING": if prop.get("distinct_count", 11) > DISTINCT_VALUE_LIMIT: example = ( f'Example: "{clean_string_values(prop["values"][0])}"' if prop["values"] else "" ) else: # If less than 10 possible values return all example = ( ( "Available options: " f'{[clean_string_values(el) for el in prop["values"]]}' ) if prop["values"] else "" ) elif prop["type"] in [ "INTEGER", "FLOAT", "DATE", "DATE_TIME", "LOCAL_DATE_TIME", ]: if prop.get("min"): # If we have min/max example = f'Min: {prop["min"]}, Max: {prop["max"]}' else: # return a single value example = ( f'Example: "{prop["values"][0]}"' if prop["values"] else "" ) elif prop["type"] == "LIST": # Skip embeddings if prop["min_size"] > LIST_LIMIT: continue example = ( f'Min Size: {prop["min_size"]}, Max Size: {prop["max_size"]}' ) formatted_rel_props.append( f" - `{prop['property']}: {prop['type']}` {example}" ) else: # Format node properties for label, props in schema["node_props"].items(): props_str = ", ".join( [f"{prop['property']}: {prop['type']}" for prop in props] ) formatted_node_props.append(f"{label} {{{props_str}}}") # Format relationship properties using structured_schema for type, props in schema["rel_props"].items(): props_str = ", ".join( [f"{prop['property']}: {prop['type']}" for prop in props] ) formatted_rel_props.append(f"{type} {{{props_str}}}") # Format relationships formatted_rels = [ f"(:{el['start']})-[:{el['type']}]->(:{el['end']})" for el in schema["relationships"] ] return "\n".join( [ "Node properties:", "\n".join(formatted_node_props), "Relationship properties:", "\n".join(formatted_rel_props), "The relationships:", "\n".join(formatted_rels), ] ) class Neo4jGraph(GraphStore): """Neo4j database wrapper for various graph operations. Parameters: url (Optional[str]): The URL of the Neo4j database server. username (Optional[str]): The username for database authentication. password (Optional[str]): The password for database authentication. database (str): The name of the database to connect to. Default is 'neo4j'. timeout (Optional[float]): The timeout for transactions in seconds. Useful for terminating long-running queries. By default, there is no timeout set. sanitize (bool): A flag to indicate whether to remove lists with more than 128 elements from results. Useful for removing embedding-like properties from database responses. Default is False. refresh_schema (bool): A flag whether to refresh schema information at initialization. Default is True. enhanced_schema (bool): A flag whether to scan the database for example values and use them in the graph schema. Default is False. driver_config (Dict): Configuration passed to Neo4j Driver. *Security note*: Make sure that the database connection uses credentials that are narrowly-scoped to only include necessary permissions. Failure to do so may result in data corruption or loss, since the calling code may attempt commands that would result in deletion, mutation of data if appropriately prompted or reading sensitive data if such data is present in the database. The best way to guard against such negative outcomes is to (as appropriate) limit the permissions granted to the credentials used with this tool. See https://python.langchain.com/docs/security for more information. """ def __init__( self, url: Optional[str] = None, username: Optional[str] = None, password: Optional[str] = None, database: Optional[str] = None, timeout: Optional[float] = None, sanitize: bool = False, refresh_schema: bool = True, *, driver_config: Optional[Dict] = None, enhanced_schema: bool = False, ) -> None: """Create a new Neo4j graph wrapper instance.""" try: import neo4j except ImportError: raise ImportError( "Could not import neo4j python package. " "Please install it with `pip install neo4j`." ) url = get_from_dict_or_env({"url": url}, "url", "NEO4J_URI") username = get_from_dict_or_env( {"username": username}, "username", "NEO4J_USERNAME" ) password = get_from_dict_or_env( {"password": password}, "password", "NEO4J_PASSWORD" ) database = get_from_dict_or_env( {"database": database}, "database", "NEO4J_DATABASE", "neo4j" ) self._driver = neo4j.GraphDatabase.driver( url, auth=(username, password), **(driver_config or {}) ) self._database = database self.timeout = timeout self.sanitize = sanitize self._enhanced_schema = enhanced_schema self.schema: str = "" self.structured_schema: Dict[str, Any] = {} # Verify connection try: self._driver.verify_connectivity() except neo4j.exceptions.ServiceUnavailable: raise ValueError( "Could not connect to Neo4j database. " "Please ensure that the url is correct" ) except neo4j.exceptions.AuthError: raise ValueError( "Could not connect to Neo4j database. " "Please ensure that the username and password are correct" ) # Set schema if refresh_schema: try: self.refresh_schema() except neo4j.exceptions.ClientError as e: if e.code == "Neo.ClientError.Procedure.ProcedureNotFound": raise ValueError( "Could not use APOC procedures. " "Please ensure the APOC plugin is installed in Neo4j and that " "'apoc.meta.data()' is allowed in Neo4j configuration " ) raise e @property def get_schema(self) -> str: """Returns the schema of the Graph""" return self.schema @property def get_structured_schema(self) -> Dict[str, Any]: """Returns the structured schema of the Graph""" return self.structured_schema def query(self, query: str, params: dict = {}) -> List[Dict[str, Any]]: """Query Neo4j database.""" from neo4j import Query from neo4j.exceptions import CypherSyntaxError with self._driver.session(database=self._database) as session: try: data = session.run(Query(text=query, timeout=self.timeout), params) json_data = [r.data() for r in data] if self.sanitize: json_data = [value_sanitize(el) for el in json_data] return json_data except CypherSyntaxError as e: raise ValueError(f"Generated Cypher Statement is not valid\n{e}") def refresh_schema(self) -> None: """ Refreshes the Neo4j graph schema information. """ from neo4j.exceptions import ClientError node_properties = [ el["output"] for el in self.query( node_properties_query, params={"EXCLUDED_LABELS": EXCLUDED_LABELS + [BASE_ENTITY_LABEL]}, ) ] rel_properties = [ el["output"] for el in self.query( rel_properties_query, params={"EXCLUDED_LABELS": EXCLUDED_RELS} ) ] relationships = [ el["output"] for el in self.query( rel_query, params={"EXCLUDED_LABELS": EXCLUDED_LABELS + [BASE_ENTITY_LABEL]}, ) ] # Get constraints & indexes try: constraint = self.query("SHOW CONSTRAINTS") index = self.query("SHOW INDEXES YIELD *") except ( ClientError ): # Read-only user might not have access to schema information constraint = [] index = [] self.structured_schema = { "node_props": {el["labels"]: el["properties"] for el in node_properties}, "rel_props": {el["type"]: el["properties"] for el in rel_properties}, "relationships": relationships, "metadata": {"constraint": constraint, "index": index}, } if self._enhanced_schema: schema_counts = self.query( "CALL apoc.meta.graphSample() YIELD nodes, relationships " "RETURN nodes, [rel in relationships | {name:apoc.any.property" "(rel, 'type'), count: apoc.any.property(rel, 'count')}]" " AS relationships" ) # Update node info for node in schema_counts[0]["nodes"]: # Skip bloom labels if node["name"] in EXCLUDED_LABELS: continue node_props = self.structured_schema["node_props"].get(node["name"]) if not node_props: # The node has no properties continue enhanced_cypher = _enhanced_schema_cypher( node["name"], node_props, node["count"] < EXHAUSTIVE_SEARCH_LIMIT ) enhanced_info = self.query(enhanced_cypher)[0]["output"] for prop in node_props: if prop["property"] in enhanced_info: prop.update(enhanced_info[prop["property"]]) # Update rel info for rel in schema_counts[0]["relationships"]: # Skip bloom labels if rel["name"] in EXCLUDED_RELS: continue rel_props = self.structured_schema["rel_props"].get(rel["name"]) if not rel_props: # The rel has no properties continue enhanced_cypher = _enhanced_schema_cypher( rel["name"], rel_props, rel["count"] < EXHAUSTIVE_SEARCH_LIMIT, is_relationship=True, ) enhanced_info = self.query(enhanced_cypher)[0]["output"] for prop in rel_props: if prop["property"] in enhanced_info: prop.update(enhanced_info[prop["property"]]) schema = _format_schema(self.structured_schema, self._enhanced_schema) self.schema = schema def add_graph_documents( self, graph_documents: List[GraphDocument], include_source: bool = False, baseEntityLabel: bool = False, ) -> None: """ This method constructs nodes and relationships in the graph based on the provided GraphDocument objects. Parameters: - graph_documents (List[GraphDocument]): A list of GraphDocument objects that contain the nodes and relationships to be added to the graph. Each GraphDocument should encapsulate the structure of part of the graph, including nodes, relationships, and the source document information. - include_source (bool, optional): If True, stores the source document and links it to nodes in the graph using the MENTIONS relationship. This is useful for tracing back the origin of data. Merges source documents based on the `id` property from the source document metadata if available; otherwise it calculates the MD5 hash of `page_content` for merging process. Defaults to False. - baseEntityLabel (bool, optional): If True, each newly created node gets a secondary __Entity__ label, which is indexed and improves import speed and performance. Defaults to False. """ if baseEntityLabel: # Check if constraint already exists constraint_exists = any( [ el["labelsOrTypes"] == [BASE_ENTITY_LABEL] and el["properties"] == ["id"] for el in self.structured_schema.get("metadata", {}).get( "constraint" ) ] ) if not constraint_exists: # Create constraint self.query( f"CREATE CONSTRAINT IF NOT EXISTS FOR (b:{BASE_ENTITY_LABEL}) " "REQUIRE b.id IS UNIQUE;" ) self.refresh_schema() # Refresh constraint information node_import_query = _get_node_import_query(baseEntityLabel, include_source) rel_import_query = _get_rel_import_query(baseEntityLabel) for document in graph_documents: if not document.source.metadata.get("id"): document.source.metadata["id"] = md5( document.source.page_content.encode("utf-8") ).hexdigest() # Import nodes self.query( node_import_query, { "data": [el.__dict__ for el in document.nodes], "document": document.source.__dict__, }, ) # Import relationships self.query( rel_import_query, { "data": [ { "source": el.source.id, "source_label": el.source.type, "target": el.target.id, "target_label": el.target.type, "type": el.type.replace(" ", "_").upper(), "properties": el.properties, } for el in document.relationships ] }, )