TripletRetriever module

TripletRetriever

Bases: Module

Retrieve triplets using a hybrid neuro-symbolic approach.

Unlike the Text2Cypher approach, this retriever is 100% guaranteed to generate a valid Cypher query every time.

It doesn't need to have the graph schema in the prompt, thus helping the language models by avoiding prompt confusion, because the nodes and relation labels are enforced dynamically using constrained structured output (similar to the Decision module).

It works by using the language model to infer the subject, object and relation labels to search for, along with a similarity search field for the object and subject triplets.

These parameters are then used to programmatically create a valid Cypher query.

This approach not only ensures the syntactical correctness of the Cypher query but also protects the graph database from Cypher injections that could arise from an adversarial prompt injection.

import synalinks
import asyncio
from typing import Literal

class Query(synalinks.DataModel):
    query: str = synalinks.Field(
        description="The user query",
    )

class Answer(synalinks.DataModel):
    query: str = synalinks.Field(
        description="The answer to the user query",
    )

class Country(synalinks.Entity):
    label: Literal["Country"]
    name: str = synalinks.Field(
        description="The country's name",
    )

class City(synalinks.Entity):
    label: Literal["City"]
    name: str = synalinks.Field(
        description="The city's name",
    )

class IsCapitalOf(synalinks.Relation):
    subj: City
    label: Literal["IsCapitalOf"]
    obj: Country

class IsCityOf(synalinks.Relation):
    subj: City
    label: Literal["IsCityOf"]
    obj: Country

language_model = synalinks.LanguageModel(
    model="ollama/mistral",
)

embedding_model = synalinks.EmbeddingModel(
    model="ollama/mxbai-embed-large"
)

knowledge_base = synalinks.KnowledgeBase(
    uri="neo4j://localhost:7687",
    entity_models=[Country, City],
    relation_models=[IsCapitalOf, IsCityOf],
    embedding_model=embedding_model,
    metric="cosine",
)

async def main():
    inputs = synalinks.Input(data_model=Query)
    x = await synalinks.KnowledgeRetriever(
        entity_models=[Country, City],
        relation_models=[IsCityOf, IsCapitalOf]
        language_model=language_model,
        knowledge_base=knowledge_base,
    )(inputs)
    outputs = await synalinks.Generator(
        data_model=Answer,
        language_model=language_model,
    )(x)

    program = synalinks.Program(
        inputs=inputs,
        outputs=outputs,
        name="kag_program",
        description="A simple KAG program",
    )

if __name__ == "__main__":
    asyncio.run(main())

Parameters:

Name	Type	Description	Default
knowledge_base	KnowledgeBase	The knowledge base to use.	None
language_model	LanguageModel	The language model to use.	None
entity_models	list	The list of entities models to search for being a list of Entity data models.	None
relation_models	list	The list of relations models to seach for. being a list of Relation data models.	None
k	int	Maximum number of similar entities to return (Defaults to 10).	10
threshold	float	Minimum similarity score for results. Entities with similarity below this threshold are excluded. Should be between 0.0 and 1.0 (Defaults to 0.5).	0.5
prompt_template	str	The default jinja2 prompt template to use (see Generator).	None
examples	list	The default list of examples, the examples are a list of tuples containing input/output JSON pairs.	None
instructions	str	The default instructions being a string containing instructions for the language model.	None
seed_instructions	list	Optional. A list of instructions to use as seed for the optimization. If not provided, use the default instructions as seed.	None
temperature	float	Optional. The temperature for the LM call.	0.0
use_inputs_schema	bool	Optional. Whether or not use the inputs schema in the prompt (Default to False) (see Generator).	False
use_outputs_schema	bool	Optional. Whether or not use the outputs schema in the prompt (Default to False) (see Generator).	False
return_inputs	bool	Optional. Whether or not to concatenate the inputs to the outputs (Default to True).	True
return_query	bool	Optional. Whether or not to concatenate the search query to the outputs (Default to True).	True
name	str	Optional. The name of the module.	None
description	str	Optional. The description of the module.	None
trainable	bool	Whether the module's variables should be trainable.	True

Source code in synalinks/src/modules/retrievers/triplet_retriever.py

@synalinks_export(
    [
        "synalinks.modules.TripletRetriever",
        "synalinks.TripletRetriever",
    ]
)
class TripletRetriever(Module):
    """Retrieve triplets using a hybrid neuro-symbolic approach.

    Unlike the Text2Cypher approach, this retriever is 100%
    guaranteed to generate a valid Cypher query **every time**.

    It doesn't need to have the graph schema in the prompt, thus
    helping the language models by avoiding prompt confusion, because
    the nodes and relation labels are enforced dynamically
    using **constrained structured output** (similar to the `Decision` module).

    It works by using the language model to infer the subject, object and
    relation labels to search for, along with a similarity search
    field for the object and subject triplets.

    These parameters are then used to *programmatically create a valid Cypher query*.

    This approach not only ensures the syntactical correctness of the
    Cypher query but also protects the graph database from Cypher
    injections that could arise from an adversarial prompt injection.

    ```python
    import synalinks
    import asyncio
    from typing import Literal

    class Query(synalinks.DataModel):
        query: str = synalinks.Field(
            description="The user query",
        )

    class Answer(synalinks.DataModel):
        query: str = synalinks.Field(
            description="The answer to the user query",
        )

    class Country(synalinks.Entity):
        label: Literal["Country"]
        name: str = synalinks.Field(
            description="The country's name",
        )

    class City(synalinks.Entity):
        label: Literal["City"]
        name: str = synalinks.Field(
            description="The city's name",
        )

    class IsCapitalOf(synalinks.Relation):
        subj: City
        label: Literal["IsCapitalOf"]
        obj: Country

    class IsCityOf(synalinks.Relation):
        subj: City
        label: Literal["IsCityOf"]
        obj: Country

    language_model = synalinks.LanguageModel(
        model="ollama/mistral",
    )

    embedding_model = synalinks.EmbeddingModel(
        model="ollama/mxbai-embed-large"
    )

    knowledge_base = synalinks.KnowledgeBase(
        uri="neo4j://localhost:7687",
        entity_models=[Country, City],
        relation_models=[IsCapitalOf, IsCityOf],
        embedding_model=embedding_model,
        metric="cosine",
    )

    async def main():
        inputs = synalinks.Input(data_model=Query)
        x = await synalinks.KnowledgeRetriever(
            entity_models=[Country, City],
            relation_models=[IsCityOf, IsCapitalOf]
            language_model=language_model,
            knowledge_base=knowledge_base,
        )(inputs)
        outputs = await synalinks.Generator(
            data_model=Answer,
            language_model=language_model,
        )(x)

        program = synalinks.Program(
            inputs=inputs,
            outputs=outputs,
            name="kag_program",
            description="A simple KAG program",
        )

    if __name__ == "__main__":
        asyncio.run(main())
    ```

    Args:
        knowledge_base (KnowledgeBase): The knowledge base to use.
        language_model (LanguageModel): The language model to use.
        entity_models (list): The list of entities models to search for
            being a list of `Entity` data models.
        relation_models (list): The list of relations models to seach for.
            being a list of `Relation` data models.
        k (int): Maximum number of similar entities to return
            (Defaults to 10).
        threshold (float): Minimum similarity score for results.
            Entities with similarity below this threshold are excluded.
            Should be between 0.0 and 1.0 (Defaults to 0.5).
        prompt_template (str): The default jinja2 prompt template
            to use (see `Generator`).
        examples (list): The default list of examples, the examples
            are a list of tuples containing input/output JSON pairs.
        instructions (str): The default instructions being a string containing
            instructions for the language model.
        seed_instructions (list): Optional. A list of instructions to use as seed for the
            optimization. If not provided, use the default instructions as seed.
        temperature (float): Optional. The temperature for the LM call.
        use_inputs_schema (bool): Optional. Whether or not use the inputs schema in
            the prompt (Default to False) (see `Generator`).
        use_outputs_schema (bool): Optional. Whether or not use the outputs schema in
            the prompt (Default to False) (see `Generator`).
        return_inputs (bool): Optional. Whether or not to concatenate the inputs to
            the outputs (Default to True).
        return_query (bool): Optional. Whether or not to concatenate the search query to
            the outputs (Default to True).
        name (str): Optional. The name of the module.
        description (str): Optional. The description of the module.
        trainable (bool): Whether the module's variables should be trainable.
    """

    def __init__(
        self,
        knowledge_base=None,
        language_model=None,
        entity_models=None,
        relation_models=None,
        k=10,
        threshold=0.5,
        prompt_template=None,
        examples=None,
        instructions=None,
        seed_instructions=None,
        temperature=0.0,
        use_inputs_schema=False,
        use_outputs_schema=False,
        return_inputs=True,
        return_query=True,
        name=None,
        description=None,
        trainable=True,
    ):
        super().__init__(
            name=name,
            description=description,
            trainable=trainable,
        )
        self.knowledge_base = knowledge_base
        self.language_model = language_model
        self.k = k
        self.threshold = threshold
        self.prompt_template = prompt_template
        self.examples = examples
        if entity_models:
            node_labels = [
                entity_model.get_schema().get("title") for entity_model in entity_models
            ]
        else:
            node_labels = []

        if relation_models:
            relation_labels = [
                relation_model.get_schema().get("title")
                for relation_model in relation_models
            ]
        else:
            relation_labels = []
        if not instructions:
            instructions = default_triplet_retriever_instructions(
                node_labels, relation_labels
            )
        self.instructions = instructions
        self.seed_instructions = seed_instructions
        self.temperature = temperature
        self.use_inputs_schema = use_inputs_schema
        self.use_outputs_schema = use_outputs_schema
        self.return_inputs = return_inputs
        self.return_query = return_query

        self.schema = TripletSearch.get_schema()

        self.schema = dynamic_enum(
            schema=self.schema,
            prop_to_update="subject_label",
            labels=node_labels,
            description="The subject label to match",
        )

        self.schema = dynamic_enum(
            schema=self.schema,
            prop_to_update="relation_label",
            labels=relation_labels,
            description="The relation label to match",
        )

        self.schema = dynamic_enum(
            schema=self.schema,
            prop_to_update="object_label",
            labels=node_labels,
            description="The object label to match",
        )

        self.query_generator = ChainOfThought(
            schema=self.schema,
            language_model=self.language_model,
            prompt_template=self.prompt_template,
            examples=self.examples,
            instructions=self.instructions,
            seed_instructions=self.seed_instructions,
            temperature=self.temperature,
            use_inputs_schema=self.use_inputs_schema,
            use_outputs_schema=self.use_outputs_schema,
            return_inputs=False,
            name="query_generator_" + self.name,
        )

    async def call(self, inputs, training=False):
        if not inputs:
            return None
        triplet_search_query = await self.query_generator(
            inputs,
            training=training,
        )
        if self.return_inputs:
            if self.return_query:
                return await ops.logical_and(
                    inputs,
                    await ops.logical_and(
                        triplet_search_query,
                        await ops.triplet_search(
                            triplet_search_query,
                            knowledge_base=self.knowledge_base,
                            k=self.k,
                            threshold=self.threshold,
                            name="similarity_search_" + self.name,
                        ),
                        name="similarity_search_with_query_and_inputs_" + self.name,
                    ),
                )
            else:
                return await ops.logical_and(
                    inputs,
                    await ops.triplet_search(
                        triplet_search_query,
                        knowledge_base=self.knowledge_base,
                        k=self.k,
                        threshold=self.threshold,
                        name="similarity_search_" + self.name,
                    ),
                    name="similarity_search_with_inputs_" + self.name,
                )
        else:
            if self.return_query:
                return await ops.logical_and(
                    triplet_search_query,
                    await ops.triplet_search(
                        triplet_search_query,
                        knowledge_base=self.knowledge_base,
                        k=self.k,
                        threshold=self.threshold,
                        name="similarity_search_" + self.name,
                    ),
                    name="similarity_search_with_query_" + self.name,
                )
            else:
                return await ops.triplet_search(
                    triplet_search_query,
                    knowledge_base=self.knowledge_base,
                    k=self.k,
                    threshold=self.threshold,
                    name="similarity_search_" + self.name,
                )

    def get_config(self):
        config = {
            "k": self.question,
            "threshold": self.labels,
            "prompt_template": self.prompt_template,
            "examples": self.examples,
            "instructions": self.instructions,
            "seed_instructions": self.seed_instructions,
            "temperature": self.temperature,
            "use_inputs_schema": self.use_inputs_schema,
            "use_outputs_schema": self.use_outputs_schema,
            "return_inputs": self.return_inputs,
            "return_query": self.return_query,
            "name": self.name,
            "description": self.description,
            "trainable": self.trainable,
        }
        knowledge_base_config = {
            "knowledge_base": serialization_lib.serialize_synalinks_object(
                self.knowledge_base,
            )
        }
        language_model_config = {
            "language_model": serialization_lib.serialize_synalinks_object(
                self.language_model,
            )
        }
        entity_models_config = {
            "entity_models": [
                (
                    serialization_lib.serialize_synalinks_object(
                        entity_model.to_symbolic_data_model(
                            name="entity_model" + (f"_{i}_" if i > 0 else "_") + self.name
                        )
                    )
                    if not is_symbolic_data_model(entity_model)
                    else serialization_lib.serialize_synalinks_object(entity_model)
                )
                for i, entity_model in enumerate(self.entity_models)
            ]
        }
        relation_models_config = {
            "relation_models": [
                (
                    serialization_lib.serialize_synalinks_object(
                        relation_model.to_symbolic_data_model(
                            name="relation_model" + (f"_{i}_" if i > 0 else "_") + self.name
                        )
                    )
                    if not is_symbolic_data_model(relation_model)
                    else serialization_lib.serialize_synalinks_object(relation_model)
                )
                for i, relation_model in enumerate(self.relation_models)
            ]
        }
        return {
            **config,
            **knowledge_base_config,
            **language_model_config,
            **entity_models_config,
            **relation_models_config,
        }

    @classmethod
    def from_config(cls, config):
        knowledge_base = serialization_lib.deserialize_synalinks_object(
            config.pop("knowledge_base"),
        )
        language_model = serialization_lib.deserialize_synalinks_object(
            config.pop("language_model"),
        )
        entity_models_config = config.pop("entity_models")
        entity_models = [
            serialization_lib.deserialize_synalinks_object(entity_model)
            for entity_model in entity_models_config
        ]
        relation_models_config = config.pop("relation_models")
        relation_models = [
            serialization_lib.deserialize_synalinks_object(relation_model)
            for relation_model in relation_models_config
        ]
        return cls(
            knowledge_base=knowledge_base,
            entity_models=entity_models,
            relation_models=relation_models,
            language_model=teacher_language_model,
            **config,
        )

default_triplet_retriever_instructions(entity_labels, relation_labels)

The default instructions for the triplet retriever

Source code in synalinks/src/modules/retrievers/triplet_retriever.py

def default_triplet_retriever_instructions(entity_labels, relation_labels):
    """The default instructions for the triplet retriever"""
    return f"""
Your task is to retrive triplets among the following entity labels: {entity_labels} and relation labels: {relation_labels}.
Think about the triplet you are looking for, which relation label do you need, then the subject and object label.
The similarity search parameters should be a short natural language string describing the entities to match.
Remember to replace the similarity search with `?` for the entity you are looking for.
""".strip()