, authorized contracts are foundational paperwork that outline the relationships, obligations, and duties between events. Whether or not it’s a partnership settlement, an NDA, or a provider contract, these paperwork usually include essential info that drives decision-making, threat administration, and compliance. Nevertheless, navigating and extracting insights from these contracts could be a complicated and time-consuming course of.
On this publish, we’ll discover how we will streamline the method of understanding and dealing with authorized contracts by implementing an end-to-end resolution utilizing Agentic Graphrag. I see GraphRAG as an umbrella time period for any technique that retrieves or causes over info saved in a data graph, enabling extra structured and context-aware responses.
By structuring authorized contracts right into a data graph in Neo4j, we will create a robust repository of knowledge that’s straightforward to question and analyze. From there, we’ll construct a LangGraph agent that permits customers to ask particular questions in regards to the contracts, making it attainable to quickly uncover new insights.
The code is accessible on this GitHub repository.
Why structuring knowledge issues
Some domains work nicely with naive RAG, however authorized contracts current distinctive challenges.
As proven within the picture, relying solely on a vector index to retrieve related chunks can introduce dangers, akin to pulling info from irrelevant contracts. It’s because authorized language is very structured, and related wording throughout totally different agreements can result in incorrect or deceptive retrieval. These limitations spotlight the necessity for a extra structured strategy, akin to GraphRAG, to make sure exact and context-aware retrieval.
To implement GraphRAG, we first must assemble a data graph.
To construct a data graph for authorized contracts, we want a strategy to extract structured info from paperwork and retailer it alongside the uncooked textual content. An LLM can assist by studying by way of contracts and figuring out key particulars akin to events, dates, contract varieties, and essential clauses. As an alternative of treating the contract as only a block of textual content, we break it down into structured parts that mirror its underlying authorized that means. For instance, an LLM can acknowledge that “ACME Inc. agrees to pay $10,000 per 30 days beginning January 1, 2024” comprises each a cost obligation and a begin date, which we will then retailer in a structured format.
As soon as we have now this structured knowledge, we retailer it in a data graph, the place entities like firms, agreements, and clauses are represented as represented together with their relationships. The unstructured textual content stays out there, however now we will use the structured layer to refine our searches and make retrieval way more exact. As an alternative of simply fetching essentially the most related textual content chunks, we will filter contracts based mostly on their attributes. This implies we will reply questions that naive RAG would wrestle with, akin to what number of contracts had been signed final month or whether or not we have now any lively agreements with a particular firm. These questions require aggregation and filtering, which isn’t attainable with commonplace vector-based retrieval alone.
By combining structured and unstructured knowledge, we additionally make retrieval extra context-aware. If a person asks a few contract’s cost phrases, we make sure that the search is constrained to the appropriate settlement slightly than counting on textual content similarity, which could pull in phrases from unrelated contracts. This hybrid strategy overcomes the constraints of naive RAG and permits for a a lot deeper and extra dependable evaluation of authorized paperwork.
Graph building
We’ll leverage an LLM to extract structured info from authorized paperwork, utilizing the CUAD (Contract Understanding Atticus Dataset), a extensively used benchmark dataset for contract evaluation licensed below CC BY 4.0. CUAD dataset comprises over 500 contracts, making it a great dataset for evaluating our structured extraction pipeline.
The token rely distribution for the contracts is visualized beneath.
Most contracts on this dataset are comparatively quick, with token counts beneath 10,000. Nevertheless, there are some for much longer contracts, with a number of reaching as much as 80,000 tokens. These lengthy contracts are uncommon, whereas shorter ones make up the bulk. The distribution reveals a steep drop-off, that means lengthy contracts are the exception slightly than the rule.
We’re utilizing Gemini-2.0-Flash for extraction, which has a 1 million token enter restrict, so dealing with these contracts isn’t an issue. Even the longest contracts in our dataset (round 80,000 tokens) match nicely inside the mannequin’s capability. Since most contracts are a lot shorter, we don’t have to fret about truncation or breaking paperwork into smaller chunks for processing.
Structured knowledge extraction
Most business LLMs have the choice to make use of Pydantic objects to outline the schema of the output. An instance for location:
class Location(BaseModel):
"""
Represents a bodily location together with deal with, metropolis, state, and nation.
"""
deal with: Optionally available[str] = Area(
..., description="The road deal with of the situation.Use None if not offered"
)
metropolis: Optionally available[str] = Area(
..., description="The town of the situation.Use None if not offered"
)
state: Optionally available[str] = Area(
..., description="The state or area of the situation.Use None if not offered"
)
nation: str = Area(
...,
description="The nation of the situation. Use the two-letter ISO commonplace.",
)When utilizing LLMs for structured output, Pydantic helps outline a transparent schema by specifying the kinds of attributes and offering descriptions that information the mannequin’s responses. Every subject has a kind, akin to str or Optionally available[str], and an outline that tells the LLM precisely format the output.
For instance, in a Location mannequin, we outline key attributes like deal with, metropolis, state, and nation, specifying what knowledge is anticipated and the way it must be structured. The nation subject, as an example, follows two-letter nation code commonplace like "US", "FR", or "JP", as an alternative of inconsistent variations like “United States” or “USA.” This precept applies to different structured knowledge as nicely, ISO 8601 retains dates in an ordinary format (YYYY-MM-DD), and so forth.
By defining structured output with Pydantic, we make LLM responses extra dependable, machine-readable, and simpler to combine into databases or APIs. Clear subject descriptions additional assist the mannequin generate accurately formatted knowledge, lowering the necessity for post-processing.
The Pydantic schema fashions may be extra refined just like the Contract mannequin beneath, which captures key particulars of a authorized settlement, making certain the extracted knowledge follows a standardized construction.
class Contract(BaseModel):
"""
Represents the important thing particulars of the contract.
"""
abstract: str = Area(
...,
description=("Excessive stage abstract of the contract with related details and particulars. Embody all related info to supply full image."
"Do no use any pronouns"),
)
contract_type: str = Area(
...,
description="The kind of contract being entered into.",
enum=CONTRACT_TYPES,
)
events: Listing[Organization] = Area(
...,
description="Listing of events concerned within the contract, with particulars of every celebration's function.",
)
effective_date: str = Area(
...,
description=(
"Enter the date when the contract turns into efficient in yyyy-MM-dd format."
"If solely the yr (e.g., 2015) is understood, use 2015-01-01 because the default date."
"At all times fill in full date"
),
)
contract_scope: str = Area(
...,
description="Description of the scope of the contract, together with rights, duties, and any limitations.",
)
length: Optionally available[str] = Area(
None,
description=(
"The length of the settlement, together with provisions for renewal or termination."
"Use ISO 8601 durations commonplace"
),
)
end_date: Optionally available[str] = Area(
None,
description=(
"The date when the contract expires. Use yyyy-MM-dd format."
"If solely the yr (e.g., 2015) is understood, use 2015-01-01 because the default date."
"At all times fill in full date"
),
)
total_amount: Optionally available[float] = Area(
None, description="Complete worth of the contract."
)
governing_law: Optionally available[Location] = Area(
None, description="The jurisdiction's legal guidelines governing the contract."
)
clauses: Optionally available[List[Clause]] = Area(
None, description=f"""Related summaries of clause varieties. Allowed clause varieties are {CLAUSE_TYPES}"""
)This contract schema organizes key particulars of authorized agreements in a structured manner, making it simpler to research with LLMs. It contains several types of clauses, akin to confidentiality or termination, every with a brief abstract. The events concerned are listed with their names, areas, and roles, whereas contract particulars cowl issues like begin and finish dates, complete worth, and governing regulation. Some attributes, akin to governing regulation, may be outlined utilizing nested fashions, enabling extra detailed and sophisticated outputs.
The nested object strategy works nicely with some AI fashions that deal with complicated knowledge relationships, whereas others might wrestle with deeply nested particulars.
We will take a look at our strategy utilizing the next instance. We’re utilizing the LangChain framework to orchestrate LLMs.
llm = ChatGoogleGenerativeAI(mannequin="gemini-2.0-flash")
llm.with_structured_output(Contract).invoke(
"Tomaz works with Neo4j since 2017 and can make a billion greenback till 2030."
"The contract was signed in Las Vegas"
)which outputs
Contract(
abstract="Tomaz works with Neo4j since 2017 and can make a billion greenback till 2030.",
contract_type="Service",
events=[
Organization(
name="Tomaz",
location=Location(
address=None,
city="Las Vegas",
state=None,
country="US"
),
role="employee"
),
Organization(
name="Neo4j",
location=Location(
address=None,
city=None,
state=None,
country="US"
),
role="employer"
)
],
effective_date="2017-01-01",
contract_scope="Tomaz will work with Neo4j",
length=None,
end_date="2030-01-01",
total_amount=1_000_000_000.0,
governing_law=None,
clauses=None
)Now that our contract knowledge is in a structured format, we will outline the Cypher question wanted to import it into Neo4j, mapping entities, relationships, and key clauses right into a graph construction. This step transforms uncooked extracted knowledge right into a queryable data graph, enabling environment friendly traversal and retrieval of contract insights.
UNWIND $knowledge AS row
MERGE (c:Contract {file_id: row.file_id})
SET c.abstract = row.abstract,
c.contract_type = row.contract_type,
c.effective_date = date(row.effective_date),
c.contract_scope = row.contract_scope,
c.length = row.length,
c.end_date = CASE WHEN row.end_date IS NOT NULL THEN date(row.end_date) ELSE NULL END,
c.total_amount = row.total_amount
WITH c, row
CALL (c, row) {
WITH c, row
WHERE row.governing_law IS NOT NULL
MERGE (c)-[:HAS_GOVERNING_LAW]->(l:Location)
SET l += row.governing_law
}
FOREACH (celebration IN row.events |
MERGE (p:Occasion {title: celebration.title})
MERGE (p)-[:HAS_LOCATION]->(pl:Location)
SET pl += celebration.location
MERGE (p)-[pr:PARTY_TO]->(c)
SET pr.function = celebration.function
)
FOREACH (clause IN row.clauses |
MERGE (c)-[:HAS_CLAUSE]->(cl:Clause {kind: clause.clause_type})
SET cl.abstract = clause.abstract
)This Cypher question imports structured contract knowledge into Neo4j by creating Contract nodes with attributes akin to abstract, contract_type, effective_date, length, and total_amount. If a governing regulation is specified, it hyperlinks the contract to a Location node. Events concerned within the contract are saved as Occasion nodes, with every celebration linked to a Location and assigned a task in relation to the contract. The question additionally processes clauses, creating Clause nodes and linking them to the contract whereas storing their kind and abstract.
After processing and importing the contracts, the ensuing graph follows the next graph schema.
Let’s additionally check out a single contract.
This graph represents a contract construction the place a contract (orange node) connects to varied clauses (crimson nodes), events (blue nodes), and areas (violet nodes). The contract has three clauses: Renewal & Termination, Legal responsibility & Indemnification, and Confidentiality & Non-Disclosure. Two events, Modus Media Worldwide and Dragon Methods, Inc., are concerned, every linked to their respective areas, Netherlands (NL) and United States (US). The contract is ruled by U.S. regulation. The contract node additionally comprises further metadata, together with dates and different related particulars.
A public read-only occasion containing CUAD authorized contracts is accessible with the next credentials.
URI: neo4j+s://demo.neo4jlabs.com
username: legalcontracts
password: legalcontracts
database: legalcontractsEntity decision
Entity decision in authorized contracts is difficult as a consequence of variations in how firms, people, and areas are referenced. An organization may seem as “Acme Inc.” in a single contract and “Acme Company” in one other, requiring a course of to find out whether or not they check with the identical entity.
One strategy is to generate candidate matches utilizing textual content embeddings or string distance metrics like Levenshtein distance. Embeddings seize semantic similarity, whereas string distance measures character-level variations. As soon as candidates are recognized, further analysis is required, evaluating metadata akin to addresses or tax IDs, analyzing shared relationships within the graph, or incorporating human evaluate for essential circumstances.
For resolving entities at scale, each open-source options like Dedupe and business instruments like Senzing supply automated strategies. Selecting the best strategy will depend on knowledge high quality, accuracy necessities, and whether or not handbook oversight is possible.
With the authorized graph constructed, we will transfer onto the agentic GraphRAG implementation.
Agentic GraphRAG
Agentic architectures fluctuate extensively in complexity, modularity, and reasoning capabilities. At their core, these architectures contain an LLM performing as a central reasoning engine, usually supplemented with instruments, reminiscence, and orchestration mechanisms. The important thing differentiator is how a lot autonomy the LLM has in making selections and the way interactions with exterior techniques are structured.
One of many easiest and simplest designs, notably for chatbot-like implementations, is a direct LLM-with-tools strategy. On this setup, the LLM serves because the decision-maker, dynamically deciding on which instruments to invoke (if any), retrying operations when crucial, and executing a number of instruments in sequence to satisfy complicated requests.
The diagram represents a easy LangGraph agent workflow. It begins at __start__, transferring to the assistant node, the place the LLM processes person enter. From there, the assistant can both name instruments to fetch related info or transition on to __end__ to finish the interplay. If a instrument is used, the assistant processes the response earlier than deciding whether or not to name one other instrument or finish the session. This construction permits the agent to autonomously decide when exterior info is required earlier than responding.
This strategy is especially well-suited to stronger business fashions like Gemini or GPT-4o, which excel at reasoning and self-correction.
Instruments
LLMs are highly effective reasoning engines, however their effectiveness usually will depend on how nicely they’re geared up with exterior instruments. These instruments , whether or not database queries, APIs, or search features, prolong an LLM’s potential to retrieve details, carry out calculations, or work together with structured knowledge.
Designing instruments which might be each normal sufficient to deal with various queries and exact sufficient to return significant outcomes is extra artwork than science. What we’re actually constructing is a semantic layer between the LLM and the underlying knowledge. Reasonably than requiring the LLM to grasp the precise construction of a Neo4j data graph or a database schema, we outline instruments that summary away these complexities.
With this strategy, the LLM doesn’t must know whether or not contract info is saved as graph nodes and relationships or as uncooked textual content in a doc retailer. It solely must invoke the appropriate instrument to fetch related knowledge based mostly on a person’s query.
In our case, the contract retrieval instrument serves as this semantic interface. When a person asks about contract phrases, obligations, or events, the LLM calls a structured question instrument that interprets the request right into a database question, retrieves related info, and presents it in a format the LLM can interpret and summarize. This allows a versatile, model-agnostic system the place totally different LLMs can work together with contract knowledge with no need direct data of its storage or construction.
There’s no one-size-fits-all commonplace for designing an optimum toolset. What works nicely for one mannequin might fail for an additional. Some fashions deal with ambiguous instrument directions gracefully, whereas others wrestle with complicated parameters or require express prompting. The trade-off between generality and task-specific effectivity means instrument design requires iteration, testing, and fine-tuning for the LLM in use.
For contract evaluation, an efficient instrument ought to retrieve contracts and summarize key phrases with out requiring customers to phrase queries rigidly. Attaining this flexibility will depend on considerate immediate engineering, sturdy schema design, and adaptation to totally different LLM capabilities. As fashions evolve, so do methods for making instruments extra intuitive and efficient.
On this part, we’ll discover totally different approaches to instrument implementation, evaluating their flexibility, effectiveness, and compatibility with numerous LLMs.
My most popular strategy is to dynamically and deterministically assemble a Cypher question and execute it towards the database. This technique ensures constant and predictable question technology whereas sustaining implementation flexibility. By structuring queries this manner, we reinforce the semantic layer, permitting person inputs to be seamlessly translated into database retrievals. This retains the LLM targeted on retrieving related info slightly than understanding the underlying knowledge mannequin.
Our instrument is meant to determine related contracts, so we have to present the LLM with choices to look contracts based mostly on numerous attributes. The enter description is once more offered as a Pydantic object.
class ContractInput(BaseModel):
min_effective_date: Optionally available[str] = Area(
None, description="Earliest contract efficient date (YYYY-MM-DD)"
)
max_effective_date: Optionally available[str] = Area(
None, description="Newest contract efficient date (YYYY-MM-DD)"
)
min_end_date: Optionally available[str] = Area(
None, description="Earliest contract finish date (YYYY-MM-DD)"
)
max_end_date: Optionally available[str] = Area(
None, description="Newest contract finish date (YYYY-MM-DD)"
)
contract_type: Optionally available[str] = Area(
None, description=f"Contract kind; legitimate varieties: {CONTRACT_TYPES}"
)
events: Optionally available[List[str]] = Area(
None, description="Listing of events concerned within the contract"
)
summary_search: Optionally available[str] = Area(
None, description="Examine abstract of the contract"
)
nation: Optionally available[str] = Area(
None, description="Nation the place the contract applies. Use the two-letter ISO commonplace."
)
lively: Optionally available[bool] = Area(None, description="Whether or not the contract is lively")
monetary_value: Optionally available[MonetaryValue] = Area(
None, description="The full quantity or worth of a contract"
)With LLM instruments, attributes can take numerous types relying on their goal. Some fields are easy strings, akin to contract_type and nation, which retailer single values. Others, like events, are lists of strings, permitting a number of entries (e.g., a number of entities concerned in a contract).
Past fundamental knowledge varieties, attributes can even signify complicated objects. For instance, monetary_value makes use of a MonetaryValue object, which incorporates structured knowledge akin to forex kind and the operator. Whereas attributes with nested objects supply a transparent and structured illustration of knowledge, fashions are inclined to wrestle to deal with them successfully, so we should always hold them easy.
As a part of this venture, we’re experimenting with an extra cypher_aggregation attribute, offering the LLM with higher flexibility for eventualities that require particular filtering or aggregation.
cypher_aggregation: Optionally available[str] = Area(
None,
description="""Customized Cypher assertion for superior aggregations and analytics.
This will likely be appended to the bottom question:
```
MATCH (c:Contract)
<filtering based mostly on different parameters>
WITH c, abstract, contract_type, contract_scope, effective_date, end_date, events, lively, monetary_value, contract_id, nations
<your cypher goes right here>
```
Examples:
1. Depend contracts by kind:
```
RETURN contract_type, rely(*) AS rely ORDER BY rely DESC
```
2. Calculate common contract length by kind:
```
WITH contract_type, effective_date, end_date
WHERE effective_date IS NOT NULL AND end_date IS NOT NULL
WITH contract_type, length.between(effective_date, end_date).days AS length
RETURN contract_type, avg(length) AS avg_duration ORDER BY avg_duration DESC
```
3. Calculate contracts per efficient date yr:
```
RETURN effective_date.yr AS yr, rely(*) AS rely ORDER BY yr
```
4. Counts the celebration with the best variety of lively contracts:
```
UNWIND events AS celebration
WITH celebration.title AS party_name, lively, rely(*) AS contract_count
WHERE lively = true
RETURN party_name, contract_count
ORDER BY contract_count DESC
LIMIT 1
```
"""The cypher_aggregation attribute permits LLMs to outline customized Cypher statements for superior aggregations and analytics. It extends the bottom question by appending question-specified aggregation logic, enabling versatile filtering and computation.
This characteristic helps use circumstances akin to counting contracts by kind, calculating common contract length, analyzing contract distributions over time, and figuring out key events based mostly on contract exercise. By leveraging this attribute, the LLM can dynamically generate insights tailor-made to particular analytical wants with out requiring predefined question constructions.
Whereas this flexibility is effective, it must be fastidiously evaluated, as elevated adaptability comes at the price of decreased consistency and robustness because of the added complexity of the operation.
We should clearly outline the operate’s title and outline when presenting it to the LLM. A well-structured description helps information the mannequin in utilizing the operate accurately, making certain it understands its goal, anticipated inputs, and outputs. This reduces ambiguity and improves the LLM’s potential to generate significant and dependable queries.
class ContractSearchTool(BaseTool):
title: str = "ContractSearch"
description: str = (
"helpful for when you must reply questions associated to any contracts"
)
args_schema: Kind[BaseModel] = ContractInputLastly, we have to implement a operate that processes the given inputs, constructs the corresponding Cypher assertion, and executes it effectively.
The core logic of the operate facilities on establishing the Cypher assertion. We start by matching the contract as the muse of the question.
cypher_statement = "MATCH (c:Contract) "Subsequent, we have to implement the operate that processes the enter parameters. On this instance, we primarily use attributes to filter contracts based mostly on the given standards.
Easy property filtering
For instance, the contract_type attribute is used to carry out easy node property filtering.
if contract_type:
filters.append("c.contract_type = $contract_type")
params["contract_type"] = contract_typeThis code provides a Cypher filter for contract_type whereas utilizing question parameters for values to stop question injection safety subject.
For the reason that attainable contract kind values are introduced within the attribute description
contract_type: Optionally available[str] = Area(
None, description=f"Contract kind; legitimate varieties: {CONTRACT_TYPES}"
)we don’t have to fret about mapping values from enter to legitimate contract varieties because the LLM will deal with that.
Inferred property filtering
We’re constructing instruments for an LLM to work together with a data graph, the place the instruments function an abstraction layer over structured queries. A key characteristic is the flexibility to make use of inferred properties at runtime, just like an ontology however dynamically computed.
if lively just isn't None:
operator = ">=" if lively else "<"
filters.append(f"c.end_date {operator} date()")Right here, lively acts as a runtime classification, figuring out whether or not a contract is ongoing (>= date()) or expired (< date()). This logic extends structured KG queries by computing properties solely when wanted, enabling extra versatile LLM reasoning. By dealing with logic like this inside instruments, we make sure the LLM interacts with simplified, intuitive operations, protecting it targeted on reasoning slightly than question formulation.
Neighbor filtering
Typically filtering will depend on neighboring nodes, akin to limiting outcomes to contracts involving particular events. The events attribute is an elective listing, and when offered, it ensures solely contracts linked to these entities are thought-about:
if events:
parties_filter = []
for i, celebration in enumerate(events):
party_param_name = f"party_{i}"
parties_filter.append(
f"""EXISTS {{
MATCH (c)<-[:PARTY_TO]-(celebration)
WHERE toLower(celebration.title) CONTAINS ${party_param_name}
}}"""
)
params[party_param_name] = celebration.decrease()This code filters contracts based mostly on their related events, treating the logic as AND, that means all specified circumstances should be met for a contract to be included. It iterates by way of the offered events listing and constructs a question the place every celebration situation should maintain.
For every celebration, a singular parameter title is generated to keep away from conflicts. The EXISTS clause ensures that the contract has a PARTY_TO relationship to a celebration whose title comprises the required worth. The title is transformed to lowercase to permit case-insensitive matching. Every celebration situation is added individually, imposing an implicit AND between them.
If extra complicated logic had been wanted, akin to supporting OR circumstances or permitting totally different matching standards, the enter would wish to alter. As an alternative of a easy listing of celebration names, a structured enter format specifying operators can be required.
Moreover, we might implement a party-matching technique that tolerates minor typos, bettering the person expertise by dealing with variations in spelling and formatting.
Customized operator filtering
So as to add extra flexibility, we will introduce an operator object as a nested attribute, permitting extra management over filtering logic. As an alternative of hardcoding comparisons, we outline an enumeration for operators and use it dynamically.
For instance, with financial values, a contract may have to be filtered based mostly on whether or not its complete quantity is larger than, lower than, or precisely equal to a specified worth. As an alternative of assuming a hard and fast comparability logic, we outline an enum that represents the attainable operators:
class NumberOperator(str, Enum):
EQUALS = "="
GREATER_THAN = ">"
LESS_THAN = "<"
class MonetaryValue(BaseModel):
"""The full quantity or worth of a contract"""
worth: float
operator: NumberOperator
if monetary_value:
filters.append(f"c.total_amount {monetary_value.operator.worth} $total_value")
params["total_value"] = monetary_value.worthThis strategy makes the system extra expressive. As an alternative of inflexible filtering guidelines, the instrument interface permits the LLM to specify not only a worth however the way it must be in contrast, making it simpler to deal with a broader vary of queries whereas protecting the LLM’s interplay easy and declarative.
Some LLMs wrestle with nested objects as inputs, making it tougher to deal with structured operator-based filtering. Including a between operator introduces further complexity because it requires two separate values, which may result in ambiguity in parsing and enter validation.
Min and Max attributes
To maintain issues easier, I are inclined to gravitate towards utilizing min and max attributes for dates, as this naturally helps vary filtering and makes the between logic simple.
if min_effective_date:
filters.append("c.effective_date >= date($min_effective_date)")
params["min_effective_date"] = min_effective_date
if max_effective_date:
filters.append("c.effective_date <= date($max_effective_date)")
params["max_effective_date"] = max_effective_dateThis operate filters contracts based mostly on an efficient date vary by including an elective decrease and higher certain situation when min_effective_date and max_effective_date are offered, making certain that solely contracts inside the specified date vary are included.
Semantic search
An attribute will also be used for semantic search, the place as an alternative of counting on a vector index upfront, we use a post-filtering strategy to metadata filtering. First, structured filters, like date ranges, financial values, or events, are utilized to slender down the candidate set. Then, vector search is carried out over this filtered subset to rank outcomes based mostly on semantic similarity.
if summary_search:
cypher_statement += (
"WITH c, vector.similarity.cosine(c.embedding, $embedding) "
"AS rating ORDER BY rating DESC WITH c, rating WHERE rating > 0.9 "
) # Outline a threshold restrict
params["embedding"] = embeddings.embed_query(summary_search)
else: # Else we kind by newest
cypher_statement += "WITH c ORDER BY c.effective_date DESC "This code applies semantic search when summary_search is offered by computing cosine similarity between the contract’s embedding and the question embedding, ordering outcomes by relevance, and filtering out low-scoring matches with a threshold of 0.9. In any other case, it defaults to sorting contracts by the latest effective_date.
Dynamic queries
The cypher aggregation attribute is an experiment I wished to check that offers the LLM a level of partial text2cypher functionality, permitting it to dynamically generate aggregations after the preliminary structured filtering. As an alternative of predefining each attainable aggregation, this strategy lets the LLM specify calculations like counts, averages, or grouped summaries on demand, making queries extra versatile and expressive. Nevertheless, since this shifts extra question logic to the LLM, making certain all generated queries work accurately turns into difficult, as malformed or incompatible Cypher statements can break execution. This trade-off between flexibility and reliability is a key consideration in designing the system.
if cypher_aggregation:
cypher_statement += """WITH c, c.abstract AS abstract, c.contract_type AS contract_type,
c.contract_scope AS contract_scope, c.effective_date AS effective_date, c.end_date AS end_date,
[(c)<-[r:PARTY_TO]-(celebration) | {celebration: celebration.title, function: r.function}] AS events, c.end_date >= date() AS lively, c.total_amount as monetary_value, c.file_id AS contract_id,
apoc.coll.toSet([(c)<-[:PARTY_TO]-(celebration)-[:LOCATED_IN]->(nation) | nation.title]) AS nations """
cypher_statement += cypher_aggregationIf no cypher aggregation is offered, we return the entire rely of recognized contracts together with solely 5 instance contracts to keep away from overwhelming the immediate. Dealing with extreme rows is essential, as an LLM scuffling with an enormous outcome set isn’t helpful. Moreover, LLM producing solutions with 100 contract titles isn’t person expertise both.
cypher_statement += """WITH acquire(c) AS nodes
RETURN {
total_count_of_contracts: measurement(nodes),
example_values: [
el in nodes[..5] |
{abstract:el.abstract, contract_type:el.contract_type,
contract_scope: el.contract_scope, file_id: el.file_id,
effective_date: el.effective_date, end_date: el.end_date,
monetary_value: el.total_amount, contract_id: el.file_id,
events: [(el)<-[r:PARTY_TO]-(celebration) | {title: celebration.title, function: r.function}],
nations: apoc.coll.toSet([(el)<-[:PARTY_TO]-()-[:LOCATED_IN]->(nation) | nation.title])}
]
} AS output"""This cypher assertion collects all matching contracts into an inventory, returning the entire rely and as much as 5 instance contracts with key attributes, together with abstract, kind, scope, dates, financial worth, related events with roles, and distinctive nation areas.
Now that our contract search instrument is constructed, we hand it off to the LLM and identical to that, we have now agentic GraphRAG carried out.
Agent Benchmark
Should you’re severe about implementing agentic GraphRAG, you want an analysis dataset, not simply as a benchmark however as a basis for all the venture. A well-constructed dataset helps outline the scope of what the system ought to deal with, making certain that preliminary improvement aligns with real-world use circumstances. Past that, it turns into a useful instrument for evaluating efficiency, permitting you to measure how nicely the LLM interacts with the graph, retrieves info, and applies reasoning. It’s additionally important for immediate engineering optimizations, letting you iteratively refine queries, instrument use, and response formatting with clear suggestions slightly than guesswork. With no structured dataset, you’re flying blind, making enhancements tougher to quantify and inconsistencies tougher to catch.
The code for the benchmark is out there on GitHub.
I’ve compiled an inventory of twenty-two questions which we’ll use to guage the system. Moreover, we’re going to introduce a brand new metric known as answer_satisfaction the place we will likely be present a customized immediate.
answer_satisfaction = AspectCritic(
title="answer_satisfaction",
definition="""You'll consider an ANSWER to a authorized QUESTION based mostly on a offered SOLUTION.
Charge the reply on a scale from 0 to 1, the place:
- 0 = incorrect, considerably incomplete, or deceptive
- 1 = right and sufficiently full
Take into account these analysis standards:
1. Factual correctness is paramount - the reply should not contradict the answer
2. The reply should deal with the core parts of the answer
3. Further related info past the answer is suitable and should improve the reply
4. Technical authorized terminology must be used appropriately if current within the resolution
5. For quantitative authorized analyses, correct figures should be offered
+ fewshots
"""Many questions can return a considerable amount of info. For instance, asking for contracts signed earlier than 2020 may yield a whole lot of outcomes. For the reason that LLM receives each the entire rely and some instance entries, our analysis ought to deal with the entire rely, slightly than which particular examples the LLM chooses to indicate.
The offered outcomes point out that each one evaluated fashions (Gemini 1.5 Professional, Gemini 2.0 Flash, and GPT-4o) carry out equally nicely for many instrument calls, with GPT-4o barely outperforming the Gemini fashions (0.82 vs. 0.77). The noticeable distinction emerges primarily when partial text2cypher is used, notably for numerous aggregation operations.
Word that that is solely 22 pretty easy questions, so we didn’t actually discover reasoning capabilities of LLMs.
Moreover, I’ve seen tasks the place accuracy may be improved considerably by leveraging Python for aggregations, as LLMs sometimes deal with Python code technology and execution higher than producing complicated Cypher queries straight.
Net Software
We’ve additionally constructed a easy React net utility, powered by LangGraph hosted on FastAPI, which streams responses on to the frontend. Particular due to Anej Gorkic for creating the online app.
You possibly can launch all the stack with the next command:
docker compose upAnd navigate to localhost:5173
Abstract
As LLMs acquire stronger reasoning capabilities, they, when paired with the appropriate instruments, can turn into highly effective brokers for navigating complicated domains like authorized contracts. On this publish, we’ve solely scratched the floor, specializing in core contract attributes whereas barely touching the wealthy number of clauses present in real-world agreements. There’s important room for progress, from increasing clause protection to refining instrument design and interplay methods.
The code is accessible on GitHub.
Pictures
All pictures on this publish had been created by the creator.