Computer-aided system for state-based, deterministic graph-controlled processing of unstructured text data under resource-efficient control of a machine language model.
The system addresses inefficiencies in unstructured text data processing by employing a persistent state data store and directed acyclic dependency structure for deterministic control, achieving reproducible and efficient text data processing across domains.
Patent Information
- Authority / Receiving Office
- DE · DE
- Patent Type
- Utility models
- Current Assignee / Owner
- GÜLDÜ, EDGAR DR
- Filing Date
- 2026-02-26
- Publication Date
- 2026-05-07
AI Technical Summary
Existing systems for processing unstructured text data using machine language models face challenges in reproducibility, resource inefficiency, and lack of deterministic control due to excessive context data transfer and redundant processing, lacking selective state-based context management.
A computer-based system with a persistent state data store, directed acyclic dependency structure, and node-specific context generation for machine language model calls, ensuring deterministic processing and reduced resource consumption by limiting data transfer and maintaining independent state-based contexts.
Enables reproducible and resource-efficient processing of unstructured text data with reduced storage and computational demands, allowing cross-domain usability without core component changes.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Technical field
[0001] The invention relates to electronic data processing, in particular systems for processing large amounts of unstructured text data using machine language models.
[0002] The invention relates in particular to a system architecture for the reproducible and resource-efficient control of such models by means of an external, deterministic dependency structure and a persistent state data storage. State of the art
[0003] Machine language models are used to analyze and generate texts. Typically, processing relies on large sets of context data, which are passed entirely to the model. This leads to increased storage requirements, high data transfer rates, and increased computational effort.
[0004] While known workflow or orchestration systems enable deterministic process control, they do not provide for the integration of machine language models with state-based context reduction.
[0005] While known graph-based orchestration systems allow for the sequential activation of processing nodes, they do not provide a node-specific selection rule that limits the context passed to a machine language model to a defined subset of the state data. In particular, such systems do not selectively transfer the context based on stored state data, but rather manage it within the framework of the respective system.
[0006] Agent-based systems often leave the control of processing to the language model itself. This results in dynamic processing paths that are not deterministic and whose results are only partially reproducible.
[0007] Furthermore, processing often takes place without a persistent structured state data store, meaning that information has to be extracted multiple times or processed redundantly.
[0008] Furthermore, well-known systems often use a continuous global context history, which increases the amount of data to be passed to the language model during processing. Object of the invention
[0009] The object of the invention is to provide a data processing system that enables reproducible processing of unstructured text data, reduces the context scope required for calls to a machine language model, decreases storage requirements, data transmission and computational effort, and provides deterministic control of the processing independent of the language model. Solution
[0010] The task is solved by a computer-based system with a persistent state data store, a directed acyclic dependency structure for deterministic control of processing, a selection rule for determining a node-specific subset of the state data for each processing node, and a state-based generation of node-specific context data for individual calls to a machine language model.
[0011] The selection rule is designed such that for each processing node, only the state data required for the respective processing function is selected. A global transfer of all previously generated data to the machine language model is avoided.
[0012] This selective context generation limits the amount of data to be transferred to the machine language model. This reduces storage requirements, data transfer, and processing time.
[0013] The processing is controlled exclusively by the external dependency structure and independently of the machine language model's output. Identical state data and identical selection rules result in identical model calls.
[0014] The system does not maintain a continuous global context history. Instead, an independent, state-based context is created for each processing node.
[0015] The dependency structure exists as an interchangeable, externally stored configuration file and is loaded at runtime. This allows the system to be used for different processing domains without changing its core components.
[0016] The system runs on at least one data processing system with a processor, main memory and persistent data storage. Example of implementation
[0017] Text data is received via an input unit (10). An extraction module (20) uses a machine language model to generate structured facts from this data. These facts are stored as state data in a persistent data storage (30).
[0018] A processing unit (40) accesses a stored directed acyclic dependency structure with a plurality of processing nodes and activates the processing nodes deterministically along the specified structure.
[0019] For each processing node, a subset of the state data is automatically generated as a node-specific context by applying selection rule (50). Selection rule (50) limits the size of the context to a predefined maximum data set.
[0020] The machine language model (60) is invoked with this node-specific context. The generated results are stored as additional state data and are available to subsequent processing nodes.
[0021] For each processing node, an execution log (70) is stored in the persistent data storage (30). The execution log enables the subsequent traceability and reproducibility of the processing.
[0022] The processing order is entirely determined by the dependency structure and occurs independently of the machine language model's output. A separate context is created for each processing node, independent of any previous model calls.
[0023] The dependency structure is stored externally as a configuration file and loaded at runtime. By replacing the configuration file, the system can be used for different processing domains without requiring changes to the core components.
[0024] This enables reproducible processing without a global context history, with reduced resource consumption and cross-domain usability. Brief description of the drawing Fig. Figure 1 shows a schematic representation of the computer-aided data processing system. The reference symbols denote: input unit (10), extraction module (20), persistent data storage (30), processing unit (40) with directed acyclic dependency structure (DAG), selection rule (50), machine language model (60), execution protocol (70), analysis result (80).
Claims
[1] Computer-based data processing system, comprising an input unit for receiving text data, Optionally, an extraction module for generating structured facts from the text data using a machine language model, or an interface for receiving externally generated structured facts. a persistent data storage for storing the generated structured facts as state data, a processing unit with a stored directed, acyclic dependency structure with a plurality of processing nodes and a selection rule for determining a node-specific subset of the state data, where the processing unit is set up, to deterministically activate the processing nodes along the dependency structure, to automatically generate a subset of the state data as a node-specific context for each processing node by applying the selection rule, to call the machine language model with the node-specific context, to store a result generated by the machine language model as additional state data in the data storage and to make it available to subsequent processing nodes, where the order of activation of the processing nodes is completely determined by the dependency structure and is independent of outputs of the machine language model, and an independent context is created for each processing node, independent of previous model calls. [2] System according to claim 1, characterized by , that the selection rule limits the size of the node-specific context to a predefined maximum data set. [3] System according to claim 1 or 2, characterized by, that an execution log is stored in persistent data storage for each processing node. [4] System according to any one of claims 1 to 3, characterized by , that the dependency structure exists as an interchangeable, externally stored configuration file and is loaded at runtime, so that the system can be used for different processing domains without changing its core components.