Vector Databases, Generative AI and Actionable Intelligence

A few years ago when we founded Numorpho Cybernetic Systems (NUMO), our goal was to create a cybernetic framework – an aggregation of multi-modal data types that could be ingested, stored, analyzed and rendered to enable what we called actionable intelligence – a composite control system that accounted for mechanisms, feedback loops, reactions to stimuli, predictive behavior and expert analysis to synthesize man-machine interactions. We divided the domain into schemas for write, read, automate, demand, evolve and measure.

Fast forward to today we are reviewing what innovative companies like KX are doing in the domain of generative AI, LLMs and time-series data management for IoT to enable our construct for Actionable Intelligence and to provide the synthetic bind between physical and digital. Called vector databases, they are designed to efficiently store, query, and analyze large volumes of time-stamped or time-series data. They are particularly suited for use cases where data points are associated with timestamps, such as sensor data, financial market data, log files, and IoT data.


Utilizing this technology will enable us to seamlessly engineer data from source to storage and all points in between to enable process automation. This will help us in the structured analysis of data as we progress it to information, knowledge, intelligence and wisdom.

Our goal is to have a comprehensive cybernetic fabric from source to storage so that node, edge, fog and cloud processing can be facilitated with secure and appropriate data available at any layer.



Cybernetics Framework

In the above whitepaper, we have defined a framework to architect the future based on a flexible model consisting of philosophical discourses (theories of awareness), existing technologies and anticipated advances in fields like AI, Genetic Programming (GP) and Quantum Computing (QC).

We have then aggregated these technologies together in a holistic fabric so that implications from one to the other can be better understood. We have also provided a way to solution complex systems pragmatically utilizing a themed approach using Systems Engineering and Design Thinking. Using the grounding principles of Computational Cybernetics, we presume that goal-oriented systems created will live within the confines of their prime directive, whilst effectively collaborating with the ecosystem. We hope this would enable better coordination between the different disciplines that are charting our future. Such future can be given thoughtful direction monitored by creative scientific minds forming a consorted effort to share information freely and not have commercial vested interests totally monopolize this emerging meta-science. Such an organization including think tanks with administrative regulatory arm and some tangible global power with ethical and social responsibility may need to become a reality, if the human civilization and its abode, the earth and ecosystems, are to survive a healthy long-term future for millions of millennia.

The designs depicted here may have a potential to provide a structure to assess the negative entropy, sustained integrated interdependent harmonious functioning, and to minimize the entropy in the larger ecosystem to achieve survival and sustenance of the most evolved earthling!


Vector databases, also known as vectorized databases or time series databases, are specialized databases designed to efficiently store, query, and analyze large volumes of time-stamped or time-series data. They are particularly suited for use cases where data points are associated with timestamps, such as sensor data, financial market data, log files, and IoT data.

A vector database makes use of generative AI to perform analytics related to similarity search as well as anomaly detection, very often making use of temporal data i.e. time-stamped data that tells us not just ‘what’ happened, but when it happened in sequence and relation to all other events in any given IT system.

Vectors are data ‘objects’, by which we mean they are numerical values that express space, place, time and various other classifying characteristics that enable us to put a granular value and meaning on our data.

Because vectors have an inherent temporal (time-aware) ability, they’re really useful for tracking how our universe of sensors are doing in the Internet of Things (IoT). As well as being extremely fast in terms of their compute power (they can ingest data and also perform actions like replication and sharding for data partitioning at high-speed), vectors enable us to build data stores that ‘understand’ the values held in different data formats in more sophisticated ways.

Traditionally and up until now, it’s been hard for a document (for example) to know much about an audio file, an image or a video – but with vector ‘embeddings’, we get a steroid boost that provides new ways to use for storage, indexing and query processing.


Vector databases provide a number of benefits that make them well-suited for time series data:

  1. Efficient storage: Vector databases optimize storage by using data compression techniques and columnar storage formats. This allows for faster data ingestion and query performance.
  2. Time-centric operations: Vector databases offer built-in support for time-centric operations, such as range queries, windowed aggregations, and time-based filtering. They provide efficient indexing and data structures to handle time-series data efficiently.
  3. Scalability: Vector databases are designed to handle high write and read loads associated with time series data. They can scale horizontally by distributing data across multiple nodes, allowing for increased storage capacity and query throughput.
  4. Analytics capabilities: Vector databases often provide integrated analytics capabilities, allowing users to perform complex analytical queries on time series data. This includes functions like interpolation, pattern matching, forecasting, and anomaly detection.
  5. Real-time processing: Vector databases excel at real-time data processing. They can handle high-frequency data updates and provide low-latency query responses, making them suitable for use cases that require real-time insights and decision-making.


Now, let’s compare vector databases with other types of databases:

  1. Graph databases: Graph databases are designed to store and analyze highly interconnected data, where relationships between entities are important. They use graph data models and specialized query languages (like Cypher) to traverse and query the graph structure efficiently. While graph databases can handle time series data, their primary focus is on relationships, whereas vector databases are optimized for time series analytics and queries.
  2. Traditional relational databases: Relational databases use structured tables with predefined schemas to store data. They are well-suited for transactional and structured data, but they may not provide optimal performance for time series data analysis. Relational databases typically require additional indexing and data modeling to handle time-based queries efficiently.
  3. Unstructured databases: Unstructured databases, such as document databases or key-value stores, are designed to handle unstructured or semi-structured data. They provide flexibility in storing and retrieving data without a rigid schema. However, they may not offer specialized features and optimizations for time series data analysis like vector databases do.
  4. NoSQL databases: NoSQL databases encompass various non-relational databases, including document stores, key-value stores, columnar databases, and more. While some NoSQL databases can handle time series data, they often lack the specific optimizations and analytics capabilities provided by vector databases. Vector databases are purpose-built for time series data and offer dedicated features for efficient storage and querying.


Vector databases can be beneficial for Large Language Models (LLMs) in several ways:

  1. Efficient storage and retrieval: LLMs generate vast amounts of text data, and storing and retrieving that data efficiently is crucial. Vector databases are designed to handle large volumes of data and provide efficient storage mechanisms optimized for time series data. This enables LLMs to store generated text in a structured and organized manner, making it easier to retrieve specific data points or ranges of data.
  2. Time-based analysis and querying: LLMs often generate text in a time-stamped manner, such as generating news articles, social media posts, or chat logs. Vector databases excel at time-centric operations, allowing LLMs to perform complex queries and analysis based on timestamps. This includes tasks like retrieving all text generated within a specific time range, aggregating text data over time intervals, or analyzing patterns and trends in the generated text over time.
  3. Real-time insights: LLMs are increasingly being used in real-time applications, such as chatbots, customer support systems, or social media analytics. Vector databases can handle high-frequency data updates and provide low-latency query responses, making them suitable for real-time processing. LLMs can leverage vector databases to store and analyze real-time generated text, enabling them to provide immediate insights and responses based on the most up-to-date information.
  4. Scalability: LLMs are resource-intensive models, and their storage and computational requirements can be substantial. Vector databases are designed to scale horizontally by distributing data across multiple nodes, allowing for increased storage capacity and query throughput. This scalability feature helps LLMs handle large amounts of generated text data and support growing workloads efficiently.
  5. Analytics capabilities: Vector databases often provide integrated analytics capabilities, such as interpolation, pattern matching, forecasting, and anomaly detection. These features enable LLMs to perform advanced analysis on the generated text, extract insights, identify patterns, and detect anomalies. LLMs can leverage the analytics capabilities of vector databases to enhance their understanding of the generated text and provide more intelligent and context-aware responses.

By utilizing vector databases, LLMs can effectively manage, analyze, and derive valuable insights from the vast amounts of text data they generate. These databases provide the necessary infrastructure and tools to handle time series data efficiently, enabling LLMs to deliver better performance, real-time insights, and enhanced capabilities in various applications and use cases.


GPT-based models themselves do not inherently use vector databases. The GPT models, such as GPT-3, GPT-4, or any other variants, are language models that focus on text generation and understanding. They operate primarily in memory and do not have direct integration with specific databases.

However, when it comes to managing and storing the generated text data produced by GPT models, organizations and applications may choose to use vector databases as a storage solution. Vector databases can provide efficient storage, indexing, and querying mechanisms for time-stamped text data generated by GPT models. They can be used as a separate component within the overall architecture to handle the persistence and retrieval of the text data.

In practice, the usage of vector databases with GPT-based models can vary depending on the specific requirements and design choices of the application. Some applications may use vector databases exclusively for storing and querying the generated text, while others may employ a hybrid approach, combining vector databases with other data storage and retrieval systems to manage different types of data or leverage specific features of other databases.

Ultimately, the decision to use vector databases or any other storage solution with GPT-based models depends on factors such as the volume of generated text data, the need for efficient querying and analysis, real-time processing requirements, scalability considerations, and the overall architecture and design of the application.


When creating a multi-modal dataset that incorporates different types of data, including text, sensory information, and other modalities, coordinating and organizing the information effectively is crucial. Here are a few considerations:

  1. Time stamp: Time stamping the data can be a useful way to coordinate different modalities, especially if the data is generated or captured at different points in time. Assigning a common time stamp to each data point allows for temporal alignment and enables time-based queries and analysis. Time stamping can be particularly relevant when dealing with time-series data or when capturing data from different sensors or sources asynchronously.
  2. Data schema and metadata: Establishing a well-defined data schema and including relevant metadata can help in coordinating and organizing the multi-modal data. The schema should capture the structure and relationships between different modalities, ensuring that the data can be stored and queried efficiently. Metadata can provide additional context and descriptive information about each modality, facilitating data integration and understanding.
  3. Data fusion techniques: To effectively combine different modalities, data fusion techniques can be applied. These techniques aim to integrate and synthesize information from multiple modalities to generate a comprehensive representation of the data. Fusion methods can vary depending on the specific use case and data characteristics, including statistical approaches, machine learning models, or rule-based systems.

Regarding real-time processing and potential latency, it depends on the nature of the sensory information and the specific requirements of your application. Real-time processing often demands low-latency data ingestion, analysis, and response times. If you require real-time processing, it is important to consider the overall system architecture, the capabilities of the data storage and processing components, and the efficiency of the chosen data coordination mechanisms.

Latency can be a concern if there are delays in capturing, transmitting, and processing the sensory information. To mitigate latency issues, you can consider optimizing the data acquisition pipeline, utilizing efficient communication protocols, leveraging distributed processing frameworks, and selecting high-performance storage and query systems. Additionally, considering the scalability and parallel processing capabilities of the chosen data storage and processing solutions can help minimize latency and ensure real-time responsiveness.

It is important to assess the specific requirements of your multi-modal data application, evaluate the available technologies, and architect the system accordingly to achieve the desired coordination, real-time processing, and low latency as needed.


Use Cases for Vector DB

Our first use case is for Smart Monitoring utilizing sensors embedded into our Additively Manufactured folding helmets. Albeit there will be different variants of the helmets based on its field use – military, industrial, construction, recreation, well care and others – our first MVP is for industrial worker safety, where data aggregated from multiple helmets (Arduino Nicla sensors) will be sent via Bluetooth for processing on Edge devices and later Wi-Fi/wired transmitted via IoT gateways to the data center or cloud provisioning.

Our goal for this is the systematic and appropriate transfer of data between ingestion and the storage devices so that we can institute actionable intelligence in the feedback loop based on a multi-modal analysis of information for different manufacturing settings – safety, predictability, quality, real time response, etc. monitoring.

Utilizing Generatieve AI

We are “prototype under development” but are moving quickly with the need to implement it in conjunction with several of our partners – both sensor providers and innovation centers where there is a lot of industry personnel who visit it on a daily basis. The smart monitoring done by helmets will coordinate with other devices / smart equipment in place to provide for a holistic view for that particular use case.

This is where we plan to use Vector Databases and Generative AI.

A pertinent use case was released by Mercedes just last week. The automaker is piloting the technology in its production process ahead of a global rollout.

We have a lot of background in Aerospace and Automotive and will utilize it in our progression with commercial and DoD implementations

Database Types we will be using

We have posted the basis for our process automation platform, Mantra M5, on LinkedIn at:

which leads to a much larger summary on our website that details the four tenets of the platform, The fourth tenet, the TAU Codex Transformer is based on a multi-modal inference engine that accounts for different databases as depicted below:

So we will be using different databases from relational (Hypersonic, SQL Server, Cockroach DB ) to Vector DB and others to appropriately storage and manage data.

Trust, that helps answer your questions.


In summary, vector databases provide specialized optimizations and analytics capabilities for time series data, making them a powerful tool for analyzing large volumes of time-stamped data efficiently. They differ from graph databases, traditional relational databases, unstructured databases, and NoSQL databases in their focus on time series data and their specialized optimizations for time-centric operations.


Rust is a systems programming language known for its focus on performance, memory safety, and concurrency. It is often used to develop high-performance and efficient software systems, including databases. While Rust can be used to build vector databases or their underlying components, it is not specific to vector databases alone.

Pinecone is a managed vector database service that offers efficient storage, indexing, and querying capabilities for high-dimensional vector data. It is designed specifically for similarity search and recommendation systems, which involve processing and searching vectors representing items, documents, or other data points. Pinecone provides an API and infrastructure that handle the complexities of vector indexing and querying, enabling developers to build applications that leverage vector similarity efficiently.

In summary, Rust is a programming language that can be used to build various software systems, including vector databases. Pinecone, on the other hand, is a managed vector database service that specializes in similarity search and recommendation use cases.

NI+IN UCHIL Founder, CEO & Technical Evangelist


Leave a Reply

%d bloggers like this: