IP Library › Granted Patent US 12,475,086
Granted Patent B2
US 12,475,086 · App. 19/040,811 · Granted Nov 18, 2025

Database constraint and rule learning using large language models

Inventors: Georg Gottlob (Paola, IT); Jinsong Guo (London, GB); Davide Mario Longo (Cosenza, IT); Yordanos Nebiyou Yifru (Messina, IT)
Assignee: Unlimidata Limited
G06F16/211G06F16/215G06F16/2365G06F16/24564
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,475,086
App. No.
19/040,811
Granted
Nov 18, 2025
Kind
B2
Abstract

A method for database constraint generation, executed by at least one processor on a computing device accessing one or more large language models (LLMs), comprising retrieving data and/or metadata from a database; generating prompts by parameterizing inputs with concrete values; interacting with LLMs through these prompts to obtain and analyze responses; and performing data intelligence processing to derive natural-language descriptions of structural database elements. The method enables generating database constraints from defined classes, such as attribute-domain restrictions, intra-relational, and inter-relational constraints. Constraints include semantic, syntactic, and dependency-based types. Orchestration of constraint learning involves predefined or dynamic workflows incorporating tasks like database sampling, constraint testing, and refinement. It employs LLM-based techniques to generate candidate rules and optimize constraints through iterative testing and scoring. The method further supports counterexample identification, score aggregation, and rule evaluation to ensure robust constraint generation and refinement.

Claims (168)

1 . A method of database constraint generation executed by at least one processor on at least one computing device and accessing one or more large language models (LLMs), the method comprising:

(i) accessing a database and retrieving at least one from data and/or metadata from the database,

(ii) generating prompts to be answered by the one or more LLMs, wherein at least one prompt is generated from a parameterized prompt by replacing parameters in the parameterized prompt with concrete values,

(iii) interacting with the one or more LLMs by sending the generated prompts to the one or more LLMs, receiving prompt-answers from the one or more LLMs, and analyzing the prompt-answers,

(iv) obtaining intelligence about a nature of data and relationships in the database or in a sub-structure of the database by automatically performing data intelligence information processing for gathering verbal descriptions of structural database elements and providing the verbal descriptions to the one or more LLMs, the data intelligence information processing comprising generating one or more data-intelligence prompts of prompt types selected from a set of prompt types comprising:

(a) prompts containing retrieved data from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database or the database sub-structure,

(b) prompts that use one or more retrieved metadata items from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database, and

(c) prompts that ask for one or more natural-language descriptions of at least one from database elements and/or database schema elements, wherein the at least one from database elements and/or database schema elements, when the database is relational, comprises data relations and attributes, the one or more natural-language descriptions to be used within subsequently generated prompts, and

(v) generating, using one or more parameterized prompts, at least one database constraint, the at least one database constraint belonging to at least one constraint class from a set of constraint classes comprising:

(a) attribute-domain restrictions that restrict the possible domain values of attributes of entities, when the database is a relational database the possible domain values of attributes of a data relation, the constraint class of attribute-domain restrictions comprising constraint of constraint types comprising:

semantic attribute-domain restrictions that restrict data items or data values to be of a certain semantic category, the category being expressed in natural language, and

syntactic attribute-value constraints, that express restrictions to attribute-values, the restrictions being syntactically verifiable or checkable,

(b) when the database is relational, intra-relational constraints that express data dependencies within a data table or data relation, and

(c) when the database is relational, inter-relational data dependencies that express data dependencies between different data relations.

2 . The method of claim 1 , wherein previous prompt-answers are used to determine one or more prompts to be sent to the one or more LLMs.

3 . The method of claim 1 , wherein at least two different prompts are respectively generated from at least two different parameterized prompts by replacing parameters in the parameterized prompts with concrete values.

4 . The method of claim 1 , wherein the database is of a type from a set of database types comprising:

(i) Relational databases, which organize and store data in a tabular format with predefined schemas,

(ii) Non-relational databases,

(iii) File-Based Databases or datasets, and

(iv) Vector databases.

5 . The method of claim 1 , wherein at least one constraint is from a set of constraint types comprising: Datalog rules, Vadalog rules, Prolog rules, rules conforming to a variant or extension of Datalog or Prolog, constraints expressed in first-order logic, tuple-generating dependencies (TGDs), numerical constraints, logical formulas containing numerical predicates, database dependencies, denial constraints and constraints expressed in natural language.

6 . The method of claim 1 , wherein sending prompts to the one or more LLMs comprises jointly sending two or more of the generated prompts to the one or more LLMs in the form of a single comprehensive prompt.

7 . The method of claim 1 , wherein the one or more metadata items from the database are obtained from at least one from:

(i) a data dictionary or a similar component of the database, and

(ii) other available metadata not being part of a data dictionary, the one or more metadata items comprising textual descriptions.

8 . The method of claim 1 , wherein at least one from:

(i) a class of the syntactic attribute-value constraints comprises at least one from:

(a) range constraints comprising the restriction of numerical or lexicographic values to one or more ranges,

(b) length constraints, restricting the length of text strings or similar data items,

(c) data format constraints, that impose one or more syntactical formats to attribute values,

(d) enumerated constraints that require a data item to belong to one or more predefined sets of explicitly enumerated values, and

(e) regular expression constraints that require a data item to match the pattern defined by a regular expression, and

(ii) a class of the intra-relational constraints comprises one or more dependencies from a set of one or more types of dependencies comprising:

(a) functional dependencies (FDs),

(b) conditional functional dependencies (CFDs),

(c) key constraints,

(d) tuple-generating dependencies (TGDs) that are intra-relational,

(e) multi-valued dependencies, and

(f) join dependencies, and

(iii) a class of the inter-relational constraints comprises one or more dependencies from a set of one or more types of dependencies comprising:

(a) inter-relational TGDs,

(b) inclusion dependencies (IDs), and

(c) foreign keys, and other referential integrity constraints.

9 . The method of claim 1 , wherein at least one from:

(i) the attribute-domain restrictions further comprise geographic constraints,

(ii) a class of the intra-relational constraints comprises monotonicity constraints, the monotonicity constraints expressing monotonic relationships between the values of two or more attributes for distinct tuples or records, and comprising:

(a) monotonicity constraints which, for specific numerical attributes A and B of a data relation R, express that, whenever two distinct tuples t1 and t2 of R have non-decreasing A-values, that is, when the A-value of t1 is less or equal than the A-value of t2 (denoted by t1[A]≤t2[A]), then t1 and t2 must also have non-decreasing B-values (t1[B]≤t2[B]), and

(b) at least one from variations and/or generalizations of the monotonicity constraints, wherein, in addition to the “S” comparison operator other comparison operators comprising “<”, “>”, “=” and/or “>” are used,

(iii) the class of the intra-relational constraints comprises symmetry constraints, wherein a symmetry constraint on a binary relation R expresses that for all entities x and y, if R relates x to y, denoted by R(x,y), then R must also relate y to x, denoted by R(y,x), wherein the symmetry constraint on R is a tuple-generating dependency of the form ∀ x ∀ y (R(x,y)→R(y,x)), or, equivalently, as a rule R(x,y)→R(y,x), where the universal quantification is understood and thus omitted, or in Prolog notation as a Prolog rule R(y,x):-R(x,y),

(iv) the class of the intra-relational constraints comprises transitivity constraints, wherein a transitivity constraint on a binary relation R corresponds to the TGD R(x,y) & R(y,z)→R(x,z), and

(v) the set of constraint classes further comprises a constraint class C, and for at least one constraint type T of the constraint class C, the class C further comprises a type of conditional T-constraints, wherein for a constraint c in T, and for a condition <cond>, a corresponding conditional T-constraint is a constraint that requires that c be satisfied only on those tuples (or groups of tuples) that satisfy <cond>, the conditional T-constraints comprising conditional functional dependencies, the conditional T-constraints further comprising a class of conditional TGDs, which are conditional constraints <cond>c, where c is a constraint representable as a TGD A 1 & . . . & A n →A 0 , where A 0 . . . A m are logical atoms, and where <cond> is condition expressible as a conjunction B 1 & . . . & B m of atoms, and where the conditional T-constraint <cond>c is representable as a rule B 1 & . . . & B m & A 1 & . . . & A n →A 0 , the class of conditional TGDs comprising as special cases conditional symmetry constraints (also called “restricted symmetry constraints”) and conditional transitivity constraints (also called “restricted transitivity constraints”).

10 . The method of claim 1 , further comprising:

(i) associating scores to records of the database, where each of the scores expresses at least one from a degree of certainty and/or a fuzzy truth value, and

(ii) evaluating at least one tuple generating dependency (TGDs) over the database, where one or more body atoms of the at least one TGD match facts that correspond to, or are identified with, database records having associated scores, and

(iii) aggregating the scores by a score aggregation which, in case a TGD is applicable and in case the TGDs rule body matches one or more scored facts, assigns a score to the resulting atom, that is, to the head atom instance obtained by applying the TGD, the score aggregation being selected from a set of score aggregation methods comprising:

(a) implicit fuzzy score aggregation, wherein a conjunction in a TGD body is interpreted as a fuzzy conjunction according to some T-norm of fuzzy logic, and

(b) explicit score aggregation, wherein the scores of each matched scored fact are contained in a distinguished argument position of the fact, and wherein the aggregation of the scores of the matched scored facts is explicitly computed by the evaluation of arithmetic atoms contained in the TGDs rule body, and

(c) scaled score aggregation, wherein the scores of each matched scored fact are aggregated using at least one arithmetic formula, in which one or more scores or sub-formulas are scaled by coefficients.

11 . The method of claim 10 , wherein the score aggregating comprises scaled score aggregation, and wherein the coefficients that scale the one or more scores or sub-formulas are determined by at least one machine learning method using as dataset the database or a subset of the database.

12 . The method of claim 1 , further comprising orchestrating constraint learning, the constraint learning orchestration automatically orchestrating a plurality of steps according to one or a combination of:

(a) a predefined fixed workflow, and

(b) a dynamic workflow wherein an order and number of repetition of steps is dynamically determined at runtime and wherein execution of at least one step depends on answers from the one or more LLMSs to prompts issued in one or more previous steps,

the constraint learning orchestration initiating, monitoring, repeating where necessary, and controlling multiple steps, each step automatically carrying out one task from a set of tasks comprising:

(i) generation of sample databases, whereby at least one sample database is obtained from the database, wherein the at least one sample-database comprises selected or sampled records from the database,

(ii) database understanding,

(iii) the generating at least one database constraint, comprising at least one from:

(a) non-LLM-based constraint generation, whereby at least one from known statistical methods and/or known methods of constraint or rule learning are applied to the database or to the at least one sample-database to generate high-confidence constraints,

(b) generation of attribute-domain restrictions,

(c) when the database is a relational database or a database that is interpretable as a relational database, bag-template-based rule generation with help of LLMs, wherein at least one prompt is issued to the one or more LLMs, the at least one prompt asking for one or more rules, each rule in a form of a tuple-generating dependency (TGD), the TGD required to be of a specified maximum length and the TGD required to conform to at least one specified bag-template, the at least one bag-template imposing restrictions on the set of predicates allowed to occur as part of atoms in the rule-body of the TGD (“allowed predicates”), and, for at least one allowed predicate, the at least one bag-template restricting the number of occurrences of the at least one predicate in the rule body of the TGD, and

(d) direct generation of classical data dependencies, wherein, for at least one type of classical database dependencies, at least one prompt is issued to the one or more LLMs asking for the recognition and generation of a set of classical dependencies deemed to hold in the database,

(iv) constraint testing, wherein a generated candidate constraint is evaluated over the database or over at least one sample database, the at least one sample database obtained from the database, and wherein strength metrics of the constraint relative to the database or sample database are computed, the computed strength metrics selected from a set of strength metrics, the set comprising confidence and support,

(v) selection of candidate constraints, wherein, according to predefined criteria based on strength-metrics values, at least one from:

(a) candidate constraints are retained for further processing,

(b) candidate constraints are rejected and eliminated, and

(c) candidate constraints are submitted to constraint refinement,

(vi) constraint refinement, wherein for at least one candidate constraint, when a constraint testing according to specified strength metrics yields strength metrics values that fulfill certain predefined conditions indicating that the constraint is at least one from not sufficiently satisfied and/or is improvable, a constraint tightening is attempted by combining the at least one constraint with at least one from further restrictions, conditions, and/or constraints, thereby producing one or more new constraints, of which those meeting predefined acceptability criteria are retained as candidate constraints,

(vii) rule set testing, wherein rules in the set of selected candidate constraints are compared with other rules in the same set in order to single out redundant rules that are subsumed or implied by other rules or sets of rules in the set of selected candidate constraints, and

(viii) final constraint selection, wherein redundant rules are eliminated and a final set of rules is output or made available for further processing.

13 . The method of claim 12 , wherein at least one sample database is generated from the database by automatically picking or sampling data tuples from the database, the automatically picking or sampling data tuples from the database comprising fixing an integer distance constant d, picking a plurality of starting tuples from the database, each starting tuple belonging to a relation of the database, and, for each picked starting tuple, adding related tuples that are connected by a join-chain of length at most d to the picked starting tuple, wherein a join-chain is a chain in which two successive tuples t and t′ join, that is, for at least one attribute A of the relation t belongs to, and for at least one attribute A′ of the relation t′ belongs to, the A-value of t coincides with the A′-value of t′, formally, t [A]-t′ [A′].

14 . The method of claim 13 , wherein at least one from:

(i) the starting tuples are chosen among tuples having at least one value that occurs infrequently in the database,

(ii) starting tuples are randomly selected, and

(iii) the related tuples are constructed via a set of join chains, wherein for join chains starting with one of the starting tuples and having length i<d, and whose last element has a value v for attribute A, a set S of new tuples is added, wherein at least one from:

(a) S consists of all new tuples having v as an argument,

(b) S consists of one or more new tuples having v as an argument, and

(c) S consists of at least one new tuple selected randomly among new tuples.

15 . The method of claim 12 , wherein the bag-template based rule-generation comprises:

(i) fixing a constant k expressing a maximum rule length,

(ii) selecting or more relation names among the names of relations of the database,

(iii) for each relation name of the one or more relation names, fixing a multiplicity, the multiplicity being a small constant expressing how often the relation name or a corresponding predicate name shall appear in a rule body, the sum of the multiplicities being at most k,

(iv) optionally selecting a rule-head predicate,

(v) issuing a prompt to the one or more LLMs, the prompt referring to at least one from the database and results of previously performed database understanding, and asking the one or more LLMs to generate TGDs assumed to be true in the database, wherein each generated TGD:

(a) has in its body at most k database atoms, wherein a database atom is an atom whose predicate is a database predicate, and

(b) when a rule-head predicate has been selected, has a head atom whose predicate is the selected head predicate, and

(c) has a rule body in which the selected database names occur with the specified multiplicity in the body of the TGD, so that if the selected multiplicity of a selected relation name R is r, then there are exactly r occurrences of atoms of the form R( . . . ) in the rule body of the TGD, wherein atoms with database predicates are allowed to be negated, and wherein no further atoms except comparison atoms occur in the rule,

(vi) any TGD rules generated by the prompt and extracted from the prompt answer are recorded as rule candidates, and

(vii) the steps (ii)-(v) are repeated by varying at least one from the selected rule head predicate, the selected database names, and/or the multiplicities.

16 . The method of claim 12 , wherein the constraint testing is performed by generating one or more computer programs or database queries that evaluate the constraint over the database and return constraint strength values according to the computed strength metrics.

17 . The method of claim 16 , wherein at least one from:

(a) the computed strength metrics are support and confidence,

(b) the one or more computer programs or database queries are generated by at least one prompt, the at least one prompt submitted to the one or more LLMs, the at least one prompt being automatically constructed and containing all necessary information relevant to the generation of the one or more computer programs or database queries, the one or more prompts further containing the request to generate one or more computer programs or database queries in the specified programming language or database query language that evaluate the constraint over the database and return the constraint strength values according to the computed strength metrics, and

(c) the constraint testing is performed by generating a database query in SQL.

18 . The method of claim 12 , wherein the set of tasks further comprises an alternative constraint testing, wherein one or more constraints from a group of constraints comprising semantic attribute-domain restrictions are tested using at least one from external resources and/or web-services.

19 . The method of claim 12 , wherein two or more sample databases are used for the constraint generation and wherein candidate constraint sets are generated separately for each of the two or more sample databases, and wherein from the candidate constraint sets a single consolidated candidate set is obtained by at least one from:

(a) selecting candidate constraints that occur in many or most of the generated candidate constraint sets,

(b) picking the strongest constraints according to some constraint strength evaluation from all generated candidate constraint sets, and

(c) uniting some or all of the candidate constraint sets.

20 . The method of claim 12 , further comprising, when the database is relational, helping the generating at least one database constraint by prior identification, via automatically issued prompts, of column clusters and of at least one from relationships and/or dependencies between column clusters, wherein a column cluster is a set of closely related attributes of the database, wherein the automatically issued prompts comprise one or more column correlation discovery prompts to identify column clusters based on the relatedness of attributes, and one or more column cluster correlation discovery prompts that ask for at least one from relationships and/or dependencies between the identified column clusters, after which the identified column clusters and the at least one from relationships and/or dependencies between column clusters are used with one or more rule generation prompts for the generating at least one database constraint.

21 . The method of claim 12 , wherein the computed strength metrics comprise confidence and support with respect to the database.

22 . The method of claim 21 , wherein for at least one candidate constraint Cand the candidate constraint selection comprises the following actions, where s1, c1 and c2 are predefined threshold constants:

(i) if the support of Cand is smaller than s1, then Cand is rejected, otherwise,

(a) if the confidence of Cand is at least c2, then Cand is retained,

(b) if the confidence of Cand is at least c1 and less than c2, then Cand is submitted to constraint refinement, and is Cand is rejected when it cannot be refined.

23 . The method of claim 12 , wherein one or more of the generated database constraints are expressed as rules each comprising a rule head and a rule body, and wherein at least one of the rules is refined by at least one from:

(i) adding an atom into the rule body, thereby restricting the applicability of the rule,

(ii) adding an atom into the rule body, thereby restricting the applicability of the rule, and replacing the rule head with a new rule head better suited for situations in which the restricted rule body applies,

(iii) determining a pivot atom and replacing the at least one rule by two new rules, one new rule by adding the pivot atom into the body of the at least one rule and the other new rule by adding a literal consisting of the negated pivot atom into the body of the at least one rule, and replacing for at least one of the two new rules the rule head by a more appropriate rule head better suited for situations in which the corresponding restricted rule body applies, and

(iv) replacing the at least one rule with at least two new rules according to at least two cases, wherein for each case, the at least one rule is replaced by a new rule by inserting into the rule body of the at least one rule one or more literals representing the case, and by adapting one or more rule heads of the new rules according to the corresponding cases, and wherein each of the one or more literals is an atom or negated atom selected from a group of atoms comprising:

i. a database atom whose predicate symbol is or corresponds to a relation name from the database,

ii. a numerical or lexicographical comparison expression,

iii. an arithmetic formula expressing a condition or an assignment to a variable, and

iv. a call to an external procedure.

24 . The method of claim 12 , wherein one or more counterexamples to satisfaction of at least one of the at least one database constraints by the database are computed by one of:

(i) a direct counterexample generation via at least one counterexample prompt automatically composed and issued to the one or more LLMs, the at least one counterexample prompt comprising asking to find counterexamples to the satisfaction by the database of the at least one database constraint, or of one or more constraints of a specified constraint type and collecting counterexamples from the prompt answer or prompt answers, or

(ii) an indirect counterexample generation by at least one LLM-generated database query program, wherein the following steps are executed:

(a) automatically generating at least one computer program prompt, the at least one computer program prompt containing information about the database, the at least one constraint or at least one constraint type, the at least one computer program prompt further containing a request to compute a database query program in a specified programming language or in a specified query language, the database query program computing counterexamples to the satisfaction of the at least one constraint or for a constraint of the at least one constraint type by the database,

(b) submitting the at least one computer program prompt to the one or more LLMs and extracting the at least one database query program from the LLM answers, and

(c) the at least one database query program is issued against the database and outputs counterexamples to the satisfaction by the database of the at least one constraint or of one or more constraints of the specified constraint type.

25 . The method of claim 24 , wherein at least one from:

(a) the at least one counterexample prompt further contains a request to exclude recognized data errors from the counterexamples,

(b) the at least one computer program prompt further contains a request to generate a database query program that excludes recognized data errors from the output,

(c) the at least one computer program prompt asks to exclude recognized data errors from the counterexamples,

(d) the at least one counterexample prompt asks for counterexamples that are outliers,

(e) the at least one computer program prompt further contains the request that the counterexamples returned by the database query program be outliers,

(f) the specified query language is SQL,

(g) the specified programming language is Prolog, and

(h) for at least one specified constraint type, one or more constraints and one or more counterexamples to the satisfaction by the database of the one or more constraints of the specified constraint type are computed automatically by use of a single prompt to the one or more LLMs.

26 . The method of claim 12 , wherein the constraint refinement comprises, for at least one constraint to be refined, using counterexamples to the satisfaction of the at least one constraint to be refined by the database as part of a refinement prompt issued to the one or more LLMs, the refinement prompt asking for refinements of the at least one constraint to be refined, wherein a counterexample to the satisfaction of the at least one constraint to be refined by the database comprises a small sub-structure of the database that provides evidence that the at least one constraint to be refined is violated.

27 . A system comprising:

at least one processor to execute instructions to:

(i) access a database and retrieve at least one from data and metadata from the database,

(ii) generate prompts to be answered by one or more large language models (LLMs), wherein at least one prompt is generated from a parameterized prompt by replacing parameters in the parameterized prompt with concrete values,

(iii) interact with the one or more LLMs by sending the generated prompts to the one or more LLMs, receiving prompt-answers from the one or more LLMs, and analyzing the prompt-answers,

(iv) obtain intelligence about a nature of data and relationships in the database or in a sub-structure of the database by automatically performing data intelligence information processing for gathering verbal descriptions of structural database elements and providing the verbal descriptions to the one or more LLMs, the data intelligence information processing comprising generating one or more data intelligence prompts of prompt types selected from a set of prompt types comprising:

(a) prompts containing retrieved data from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database or the database sub-structure,

(b) prompts that use one or more retrieved metadata items from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database, and

(c) prompts that ask for one or more natural-language descriptions of at least one from data base elements and database schema elements, wherein the at least one from database elements and database schema elements, when the database is relational, comprises data relations and attributes, the one or more natural-language descriptions to be used within subsequently generated prompts,

(v) generate, using one or more parameterized prompts, at least one database constraint, the at least one database constraint belonging to at least one constraint class from a set of constraint classes comprising:

(a) attribute-domain restrictions that restrict the possible domain values of attributes of entities, when the database is a relational database the possible domain values of attributes of a data relation, the constraint class of attribute-domain restrictions comprising constraint of constraint types comprising:

semantic attribute-domain restrictions that restrict data items or data values to be of a certain semantic category, the category being expressed in natural language, and

syntactic attribute-value constraints, that express restrictions to attribute-values, the restrictions being syntactically verifiable or checkable,

(b) when the database is relational, intra-relational constraints that express data dependencies within a data table or data relation, and

(c) when the database is relational, inter-relational data dependencies that express data dependencies between different data relations.

28 . A non-transitory computer-readable storage medium, having instructions stored thereon that, when executed by a computing device cause the computing device to perform operations, comprising:

(i) accessing a database and retrieving at least one from data and/or metadata from the database,

(ii) generating prompts to be answered by the one or more LLMs, wherein at least one prompt is generated from a parameterized prompt by replacing parameters in the parameterized prompt with concrete values,

(iii) interacting with the one or more LLMs by sending the generated prompts to the one or more LLMs, receiving prompt-answers from the one or more LLMs, and analyzing the prompt-answers,

(iv) obtaining intelligence about a nature of data and relationships in the database or in a sub-structure of the database by automatically performing data intelligence information processing for gathering verbal descriptions of structural database elements and providing the verbal descriptions to the one or more LLMs, the data intelligence information processing comprising generating one or more data-intelligence prompts of prompt types selected from a set of prompt types comprising:

(a) prompts containing retrieved data from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database or the database sub-structure,

(b) prompts that use one or more retrieved metadata items from the database to inform one or more LLMs about the nature of the data and the relationships contained in the database, and

(c) prompts that ask for one or more natural-language descriptions of at least one from database elements and/or database schema elements, wherein the at least one from database elements and/or database schema elements, when the database is relational, comprises data relations and attributes, the one or more natural-language descriptions to be used within subsequently generated prompts, and

(v) generating, using one or more parameterized prompts, at least one database constraint, the at least one database constraint belonging to at least one constraint class from a set of constraint classes comprising:

(a) attribute-domain restrictions that restrict the possible domain values of attributes of entities, when the database is a relational database the possible domain values of attributes of a data relation, the constraint class of attribute-domain restrictions comprising constraint of constraint types comprising:

semantic attribute-domain restrictions that restrict data items or data values to be of a certain semantic category, the category being expressed in natural language, and

syntactic attribute-value constraints, that express restrictions to attribute-values, the restrictions being syntactically verifiable or checkable,

(b) when the database is relational, intra-relational constraints that express data dependencies within a data table or data relation, and

(c) when the database is relational, inter-relational data dependencies that express data dependencies between different data relations.

Assignments (3)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Oct 21, 2025
From: RATIOLYTICS LIMITED
To: UNLIMIDATA LIMITED
Reel/Frame 072614/0516 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Mar 26, 2025
From: RATIOLYTICS LIMITED
To: UNLIMIDATA LIMITED
Reel/Frame 070641/0391 →
NUNC PRO TUNC ASSIGNMENT Recorded Feb 6, 2025
From: GOTTLOB, GEORG; GUO, JINSONG; LONGO, DAVIDE MARIO; YIFRU, YORDANOS NEBIYOU
To: RATIOLYTICS LIMITED
Reel/Frame 070137/0640 →
Continuity (3)
Continuation 18792699 · Aug 2, 2024
Provisional Application 63530795 · Aug 4, 2023
Related Publication 20250173314A1 · May 29, 2025
References Cited (94)
US 8577849B2 · Yakout et al. · 2013 [cited by applicant]
US 11681689B2 · Seth et al. · 2023 [cited by applicant]
US 11966371B1 · Jennings · 2024 [cited by applicant]
US 20050182739A1 · Dasu et al. · 2005 [cited by applicant]
US 20160092557A1 · Stojanovic et al. · 2016 [cited by applicant]
US 20170169090A1 · Jose et al. · 2017 [cited by applicant]
US 20190005118A1 · Tripathi et al. · 2019 [cited by applicant]
US 20210173817A1 · Bates-Haus et al. · 2021 [cited by applicant]
US 20210248206A1 · Dube-Cousineau · 2021 [cited by applicant]
US 20220050838A1 · Piatetsky et al. · 2022 [cited by applicant]
US 20220092041A1 · Raack et al. · 2022 [cited by applicant]
US 20240362968A1 · Lyons · 2024 [cited by examiner]
US 20250061307A1 · Tran · 2025 [cited by examiner]
Xu, X. et al, How to Unleash the Power of Large Language Models for Few-shot Relation Extraction, Proceedings of The Fourth Workshop on Simple and Efficient Natural Language Processing, Jul. 13, 2023, 11 pages. [cited by applicant]
Fernandez, R. C. et al., How Large Language Models Will Disrupt Data Management, Proceedings of the VLDB Endowment, vol. 16, No. 11, Jul. 1, 2023, 8 pages. [cited by applicant]
Brinkmann, A. et al., Product Information Extraction using ChatGPT, arXiv, Jun. 23, 2023, 5 pages. [cited by applicant]
Peeters, R. et al., Using ChatGPT for Entity Matching, arXiv, Jul. 22, 2023, 10 pages. [cited by applicant]
Urban, M. et al., OmniscientDB: A Large Language Model-Augmented DBMS That Knows What Other DBMSs Do Not Know, Sixth International Workshop on Exploiting Artifi-cial Intelligence Techniques for Data Management, Jun. 20,… [cited by applicant]
Ilyas, I. F. et al., Growing and Serving Large Open-domain Knowledge Graphs, SIGMOD '23: Companion of the 2023 International Conference on Management of Data, Jun. 5, 2023, 7 pages. [cited by applicant]
Dong, Y. et al., DeepJoin: Joinable Table Discovery with Pre-trained Language Models, Proceedings of the VLDB Endowment, vol. 16, No. 10, Jun. 1, 2023, 13 pages. [cited by applicant]
Naeem, Z. A. et al., RetClean: Retrieval-Based Data Cleaning Using Foundation Models and Data Lakes, arXiv, Mar. 29, 2023, 5 pages. [cited by applicant]
Fang, L. et al., KAER: A Knowledge Augmented Pre-Trained Language Model for Entity Resolution, arXiv, Jan. 12, 2023, 6 pages. [cited by applicant]
Chen, Z. et al., Symphony: Towards Natural Language Query Answering over Multi-modal Data Lakes, 13th Annual Conference on Innovative Data Systems Research, Jan. 8, 2023, 7 pages. [cited by applicant]
Narayan, A. et al., Can Foundation Models Wrangle Your Data?, arXiv, Dec. 24, 2022, 12 pages. [cited by applicant]
Vos, D. et al., Towards Parameter-Efficient Automation of Data Wrangling Tasks with Prefix-Tuning, Table Representation Learning Workshop at NeurIPS 2022, Oct. 21, 2022, 9 pages. [cited by applicant]
Trummer, Immanuel, From BERT to GPT-3 Codex: Harnessing the Potential of Very Large Language Models for Data Management, Proceedings of the VLDB Endowment, vol. 15, No. 12, Aug. 1, 2022, 4 pages. [cited by applicant]
Negreanu, C. et al., Rows from Many Sources: Enriching row completions from Wikidata with a pre-trained Language Model, Companion Proceedings of the Web Conference 2022 , Aug. 16, 2022, 9 pages. [cited by applicant]
Xie, X. et al., From Discrimination to Generation: Knowledge Graph Completion with Generative Transformer, Companion Proceedings of the Web Conference 2022, Aug. 16, 2022, 4 pages. [cited by applicant]
Fan, W. et al., Parallel Rule Discovery from Large Datasets by Sampling, SIGMOD '22: Companion of the 2022 International Conference on Management of Data, Jun. 11, 2022, 15 pages. [cited by applicant]
Chen, X, et al., KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization for Relation Extraction, arXiv, Apr. 25, 2022, 11 pages. [cited by applicant]
Al Khamissi, B. et al., A Review on Language Models as Knowledge Bases, arXiv, Apr. 12, 2022, 21 pages. [cited by applicant]
Qin, G. et al., Learning How to Ask: Querying LMs with Mixtures of Soft Prompts, Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo… [cited by applicant]
Shin, T. et al., AUTOPROMPT: Eliciting Knowledge from Language Models with Automatically Generated Prompts, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, Nov. 16, 2020, 14 pages. [cited by applicant]
Petroni, F. et al., Language Models as Knowledge Bases?, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, … [cited by applicant]
Pena, E. H. M. et al., Discovery of Approximate (and Exact) Denial Constraints, Proceedings of the VLDB Endowment, vol. 13, No. 3, Nov. 1, 2019, 13 pages. [cited by applicant]
Mohamed, Y. et al., Tailoring Rule-Based Data Quality Assessment to the Patient-Centered Outcomes Research Network (PCORnet) Common Data Model (CDM), AMIA Annual Symposium Proceedings, Apr. 2023, 775-784. [cited by applicant]
Caruccio, L., et al., Discovering Functional Dependencies: Can We Use ChatGPT to Generate Algorithms?, CEUR Workshop Proceedings, Sep. 2023, 12 pages. [cited by applicant]
Sui, Y. et al., Table Meets LLM: Can Large Language Models Understand Structured Table Data? A Benchmark and Empirical Study, Proceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM '24)… [cited by applicant]
Karapantelakis, A. et al., Using Large Language Models to Understand Telecom Standards, ArXiv, Apr. 2024, 8 pages. [cited by applicant]
Tamm, H. C. et al., Towards augmented data quality management: Automation of Data Quality Rule Definition in Data Warehouses, ArXiv, Jul. 2024, 1-70. [cited by applicant]
Ghosh, S., Taming the Data Deluge: Automating Data Quality with Generative AI, Medium, May 3, 2024, 12 pages. [cited by applicant]
Vertsel, A., et al., Hybrid LLM/Rule-based Approaches to Business Insights Generation from Structured Data, ArXiv, Apr. 2024, 28 pages. [cited by applicant]
Règin, F. et al., Combining Constraint Programming Reasoning With Large Language Model Predictions, ArXiv, Jul. 2024, 15 pages. [cited by applicant]
Shankar, S., et al., spade: Synthesizing Data Quality Assertions for Large Language Model Pipelines, ArXiv, Mar. 2024, 17 pages. [cited by applicant]
Fang, X. et al., Large Language Models (LLMs) on Tabular Data: Prediction, Generation, and Understanding—A Survey, Transactions on Machine Learning Research, 2024, 46 pages. [cited by applicant]
Zhu, Y. et al., LLMs for knowledge graph construction and reasoning: recent capabilities and future opportunities, Neuro-Symbolic Intelligence: Large Language Model Enabled Knowledge Engineering, Aug. 21, 2024, 23 pages. [cited by applicant]
Feuer, B. et al., ArcheType: A Novel Framework for Open-Source col. Type Annotation using Large Language Models, ACM Digital Library, Aug. 6, 2024, 14 pages. [cited by applicant]
Remadi, A. et al., To prompt or not to prompt: Navigating the use of Large Language Models for integrating and modeling heterogeneous data, Data and Knowledge Engineering, Jul. 2024, 17 pages. [cited by applicant]
Wang, T. et al., Match, Compare, or Select? An Investigation of Large Language Models for Entity Matching, arXiv, Jun. 23, 2024, 13 pages. [cited by applicant]
Wang, S. et al., Can LLMs Reason with Rules? Logic Scaffolding for Stress-Testing and Improving LLMs, arXiv, Jun. 21, 2024, 22 pages. [cited by applicant]
Sun, K. et al., Head-to-Tail: How Knowledgeable are Large Language Models (LLMs)? A.K.A. Will LLMs Replace Knowledge Graphs?, Proceedings of the 2024 Conference of the North American Chapter of the Association for Compu… [cited by applicant]
Akella, A. et al., An Automatic Prompt Generation System for Tabular Data Tasks, Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo… [cited by applicant]
Li, X. et al., Towards Efficient Data Wrangling with LLMs using Code Generation, Workshop on Data Management for End-to-End Machine Learning, Jun. 9, 2024, 5 pages. [cited by applicant]
Zezhou, Huang, Disambiguate Entity Matching using Large Language Models through Relation Discovery, Governance, Understanding and Integration of Data for Effective and Responsible AI, Jun. 9, 2024, 4 pages. [cited by applicant]
Kayali, M. et al., CHORUS: Foundation Models for Unified Data Discovery and Exploration, arXiv, May 31, 2024, 11 pages. [cited by applicant]
Döhmen, T. et al., SchemaPile: A Large Collection of Relational Database Schemas, Proceedings of the ACM on Management of Data, Jun. 2024, 25 pages. [cited by applicant]
Sheetrit, E. et al., ReMatch: Retrieval Enhanced Schema Matching with LLMs, arXiv, May 30, 2024, 5 pages. [cited by applicant]
Qian, Y. et al., UNIDM: a Unified Framework for Data Manipulation With Large Language Models, Proceedings of the 5th MLSys Conference, May 13, 2024, 18 pages. [cited by applicant]
Zhang, Z. et al., Directions Towards Efficient and Automated Data Wrangling with Large Language Models, 2024 IEEE 40th International Conference on Data Engineering Workshops, May 13, 2024, 4 pages. [cited by applicant]
Fan, M. et al., Cost-Effective In-Context Learning for Entity Resolution: A Design Space Exploration, 2024 IEEE 40th International Conference on Data Engineering, May 13, 2024, 14 pages. [cited by applicant]
Li, H. et al., BoostER: Leveraging Large Language Models for Enhancing Entity Resolution, Companion Proceedings of the ACM Web Conference 2024, May 13, 2024, 4 pages. [cited by applicant]
Xia, Y. et al., APrompt4EM: Augmented Prompt Tuning for Generalized Entity Matching, arXiv, May 8, 2024, 14 pages. [cited by applicant]
Sufi, Fahim., Addressing Data Scarcity in the Medical Domain: A GPT-Based Approach for Synthetic Data Generation and Feature Extraction, Information, May 6, 2024, 31 pages. [cited by applicant]
Wornow, M. et al., Automating the Enterprise with Foundation Models, arXiv, May 3, 2024, 9 pages. [cited by applicant]
Gao, D. et al., Text-to-SQL Empowered by Large Language Models: A Benchmark Evaluation, Proceedings of the VLDB Endowment, vol. 17, No. 5, May 2, 2024, 14 pages. [cited by applicant]
Khalid, H. et al., Repairing raw metadata for metadata management, Information Systems, May 2024, 18 pages. [cited by applicant]
Biester, F. et al., LLMClean: Context-Aware Tabular Data Cleaning via LLM-Generated OFDs, arXiv, Apr. 29, 2024, 13 pages. [cited by applicant]
Chen, Z. et al., SEED: Domain-Specific Data Curation With Large Language Models, arXiv, Apr. 24, 2024, 20 pages. [cited by applicant]
Abdelaal, M. et al., SAGED: Few-Shot Meta Learning for Tabular Data Error Detection, Proceedings of the 27th International Conference on Extending Database Technology, Mar. 28, 2024, 13 pages. [cited by applicant]
Saeed, M. et al., Querying Large Language Models with SQL, Proceedings of the 27th International Conference on Extending Database Technology, Mar. 28, 2024, 8 pages. [cited by applicant]
Pereira, J. L. M. et al., Cleenex: Support for User Involvement during an Iterative Data Cleaning Process, ACM J. Data Inform. Quality, vol. 16, No. 1, Mar. 18, 2024, 26 pages. [cited by applicant]
Liu, J. et al., Query Rewriting via Large Language Models, arXiv, Mar. 14, 2024, 13 pages. [cited by applicant]
Bordt, S. et al., Elephants Never Forget: Testing Language Models for Memorization of Tabular Data, arXiv, Mar. 11, 2024, 24 pages. [cited by applicant]
Cong, T. et al., Observatory: Characterizing Embeddings of Relational Tables, VDLB Endowment, Mar. 5, 2024, 14 pages. [cited by applicant]
Luo, L. et al., Reasoning on Graphs: Faithful and Inter-Pretable Large Language Model Reasoning, arXiv, Feb. 24, 2024, 24 pages. [cited by applicant]
Caufield, J. H. et al., Structured Prompt Interrogation and Recursive Extraction of Semantics (SPIRES): a method for populating knowledge bases using zero-shot learning, Bioinformatics, Feb. 21, 2024, 10 pages. [cited by applicant]
Sufi, Fahim, Generative Pre-Trained Transformer (GPT) in Research: A Systematic Review on Data Augmentation, Information, Feb. 8, 2024, 27 pages. [cited by applicant]
Lin, Y. et al., SMARTFEAT: Efficient Feature Construction through Feature-Level Foundation Model Interactions, 14th Annual Conference on Innovative Data Systems Research, Jan. 14, 2024, 7 pages. [cited by applicant]
Pan, S. et al., Unifying Large Language Models and Knowledge Graphs: A Roadmap, 3580 IEEE Transactions on Knowledge and Data Engineering, vol. 36, No. 7, Jan. 10, 2024, 20 pages. [cited by applicant]
Yu, Y. et al., Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias, 37th Conference on Neural Information Processing Systems, Dec. 16, 2023, 51 pages. [cited by applicant]
Jiang, Z. et al., Pre-trained Tabular Transformer for Real-time, Efficient, Stable Radiomics Data Processing: A Comprehensive Study, 2023 IEEE International Conference on E-health Networking, Application & Services, Dec… [cited by applicant]
Huang, Y. et al., Interactive Table Synthesis With Natural Language, IEEE Transactions on Visualization and Computer Graphics, vol. 30, No. 9, Nov. 1, 2023, 15 pages. [cited by applicant]
Singh, M. et al., FormaT5: Abstention and Examples for Conditional Table Formatting with Natural Language, Proceedings of the VLDB Endowment, vol. 17, No. 3, Nov. 1, 2023, 14 pages. [cited by applicant]
Wu, S. et al., Blocker and Matcher Can Mutually Benefit: A Co-Learning Framework for Low-Resource Entity Resolution, Proceedings of the VLDB Endowment, vol. 17, No. 3, Nov. 1, 2023, 13 pages. [cited by applicant]
Shin, C. et al., Pool-Search-Demonstrate: Improving Data-wrangling LLMs via better in-context examples, Table Representation Learning Workshop at NeurIPS 2023, Oct. 28, 2023, 14 pages. [cited by applicant]
Naeim Abadi, A. et al., Product Entity Matching via Tabular Data, Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, Oct. 21, 2023, 5 pages. [cited by applicant]
Li, P. et al., Table-GPT: Table-tuned GPT for Diverse Table Tasks, arXiv, Oct. 13, 2023, 47 pages. [cited by applicant]
Luo, L. et al., ChatRule: Mining Logical Rules with Large Language Models for Knowledge Graph Reasoning, arXiv, Sep. 4, 2023, 11 pages. [cited by applicant]
Trummer, Immanuel, Demonstrating NaturalMiner: Searching Large Data Sets for Abstract Patterns Described in Natural Language, SIGMOD '23: Companion of the 2023 International Conference on Management of Data, Jun. 15, 20… [cited by applicant]
Jo, S. et al., Demonstration of ThalamusDB: Answering Complex SQL Queries with Natural Language Predicates on Multi-Modal Data, SIGMOD '23: Companion of the 2023 International Conference on Management of Data, Jun. 5, 2… [cited by applicant]
Arora, S. et al., Language Models Enable Simple Systems for Generating Structured Views of Heterogeneous Data Lakes, Proceedings of the VLDB Endowment, Apr. 19, 2023, 14 pages. [cited by applicant]
Chen, Z. et al., Lingua Manga : A Generic Large Language Model Centric System for Data Curation, arXiv, Sep. 1, 2023, 4 pages. [cited by applicant]
Zhang, H. et al., Large Language Models as Data Preprocessors, arXiv, Aug. 30, 2023, 5 pages. [cited by applicant]
Korini, K. et al., Column Type Annotation using ChatGPT, CEUR Workshop Proceedings, Aug. 28, 2023, 10 pages. [cited by applicant]