• Edited

Our AI assistant, a.k.a. codi, also deserves its own changelog thread.

Especially after the publication of flespi roadmap 2024, where we highlighted a focus on the application of AI tech in our telematics services.

Here, we accumulate changes in codi's architecture and post updates to its capabilities as soon as they are integrated.

The list with actual codi capabilities:

  • It can support you answering any flespi-related question loaded with the knowledge published in our comprehensive Knowledge Base and 7 year old Blog.
  • It can generate code for your flespi-related API calls in any language, consult on flespi REST API, its structure and parameters to use.
  • It can generate you PVM code to transform message parameters and parse binary and textual payloads upon your request. You may supply example in your code of choice and it will try to provide you with corresponding PVM code.
  • It can generate flespi expressions for you upon request.
  • It can consult you on the integrated protocols, devices and explain the meaning and protocol-specific source for any message parameter within some protocol.
  • It can show you how to control device remotely, which API calls to use and how to correctly construct command API call for your specific device.
  • It can help you to troubleshoot and diagnose what is wrong and propose changes in the configuration of devices, channels, streams, calculators, plugins and webhooks that you created.
  • It can navigate you through parameters reported by the particular device (telemetry).
  • It can assist you with your account usage and limitations or suggest how to save money on your flespi bill.
  • It can consult you on the device configuration, remote management, peripheral usage for hardware produced by Teltonika, Ruptela, Xirgo, Galileosky, Suntech, Calamp, Howen, few most popular device models by Queclink, Jimi Iot(Concox), Topflytech and some other brands. That knowledge is based on the information from official device manuals.

Recently we installed beta-version of next AI platform for the codi and now its implementation is totally different comparing to the initial. A little bit later we will publish a dedicated blog article with more insight about new codi architecture.

A set of new codi capabilities:

  • Loaded with REST API knowledge with the possibility to generate programming code on any language accessing flespi API.
  • Generate flespi expressions upon your request.
  • Provided knowledge of integrated protocols, devices and trained to understand per-protocol message parameters, their meaning and protocol-specific origin of each param, (e.g. Teltonika's AVL ID 199 and 357 mapped to segment.vehicle.mileage parameter in flespi)
    • Edited

    We enhanced codi(AI assistant) capabilities with:

    • Troubleshooting and diagnose what is wrong in your account with basic items. It now propose specific changes or explain meaning of the configuration of your devices, channels, streams, calculators, plugins and even webhooks.
    • It can navigate you through parameters reported by the particular device (telemetry).
    • And it can help you to understand which settings and commands are available and how to construct API call to remotely control some parameters of specific devices.

      We trained codi to generate PVM code on your request. With PVM code you can transform device message, parse it fields into different parameters (BLE beacons, data from serial port, etc) and in general do with message whatever you want before it is registered in flespi and streamed to your servers.
      Maybe not all tasks, but quite capable.

      You can supply to it either description of the task or just a code block on your favorite programming language and it will try to provide you corresponding pvm code. PVM code is not yet automatically validated and may be incorrect, but you may provide codi feedback what is wrong and it will try to fix it. Most of popular tasks it should be able to solve already.


        25 days later

        This week we installed 3rd version of our AI assistant platform. Principal change is that we got rid of specific QnA experts and now using two-stage processing to determine correct context of the discussion. And based on the context we are preloading different knowledge, instructions for LLM and even choosing different LLM/model for answer preparation based on the user capabilities. One model for a bla-bla-bla discussion, another model to solve deep technical problems, third model for sales consulting and so on ;-)

        As a bonus for our users now codi has a knowledge about all protocols and devices integrated into flespi. It can search and propose you OBD, CAN, personal trackers, pet trackers, assets trackers and so on. And even recommend some settings to adjust in your devices:

        We are still in the very early phase for this but within few months we will have enough knowledge to cover top-5 or even top-10 device manufacturers.

          a month later

          We recharged codi with a knowledge about user account limitations and counters, including billing information.
          Now it is able to consult and advice some actions on how to save money in your flespi account ;-) Enjoy!

          P.S. Be aware that codi's math skills are not excellent and sometimes it multiply numbers incorrectly. For small numbers, like those that you may have in you invoice it happens in rare cases. However if take into account big numbers like UNIX timestamp - without calculation tools it hallucinates on its own quite a lot.

            Remember in our roadmap 2024 I wrote about the attempt to enrich codi's knowledge with device-specific?

            Now it is able to consult you about Teltonika devices. Fully loaded with all official documentation it now knows a lot about each Teltonika device model: commands, parameters, configuration, firmware, FAQs, basically everything that you can find on the Teltonika Wiki.



            When using this feature make sure you provided into chat context either information about your flespi device ID or at least specify device model and manufacturer, e.g. Teltonika FMB920. If you will give your device ID codi will load the knowledge for its particular model.

            And yes, this feature is very experimental and we will adopt it based on real life scenarios and user expectations. Do not push codi too hard.

            a month later

            We enriched codi with the knowledge from official device manuals by: Ruptela, Xirgo and Galileosky.

            Now every request from you with indication of device ID or device type name loads into AI context also quite a lot of information from manufacturer's manual.

            Also we enabled quick AI assistance using manufacturer's manual directly in the device. This functionality is limited only to the information from manual and do not have access to device state, but can be used as a quick and efficient way to navigate through device manual.

              7 days later

              We loaded codi with the knowledge extracted from Howen Technologies manuals for most popular devices. Now it can effectively consult you how to configure, operate, troubleshoot and install Howen video devices.

                • Edited

                We are now providing the AI assistant service for most popular Jimi IoT, Suntech and Topflytech devices.
                Also for one manufacturer - Calamp, which recently filed bankruptcy we integrated into the device AI assistant the knowledge on writing PEG scripts (special calamp scripts for controlling device logic) and other general Calamp-specific information from its manuals.

                  9 days later
                  • Edited

                  AI assistance is available for Atrack devices. AI assistent ready to help with the reports for all the models. AK11, AK500 and AS500 also equipped with device Manuals knowledge.

                  Reporting protocol assistance is also avialble for Streamax devices and the group of 3 scooter and bike sharing related protocols: Omni, Freego and Segway-Ninebot

                    14 days later
                    22 days later

                    Based on real-life users expectations that clicked on "AI" tab in device and asked questions relative to flespi platform or specifically this device data we modified knowledge sources for that assistant and empowered it with flespi knowledge.

                    Now you may chat with this AI assistant in fully unguided mode and it will support you both in device configuration and in flespi in general. However it still lacks access to the real data from device. To uncover the full AI power we suggest to use CHAT button in top-right corner of your flespi.io account and communicate with codi.

                    For the API users of this consultation service it will continue to operate in pure device documentation mode by default.

                      3 months later

                      Step by step codi's architecture evolved into gen3.1. Not that much due to the fundamental changes but we implemented hundreds of minor improvements based on the practical usage. And raised codi's intelligence to a next level.

                      The most notable change is the model - codi is now mostly powered by the latest and most-intelligent model from Anthropic - claude-3-5-sonnet-20241022. After we improved knowledge retrieval system it's capabilities in resolving the majority of issues are now in par with human engineers. Just give it a try! You can find codi in top-right corner of your flespi.io account under CHAT button.

                      And another AI tool by flespi (open source) for your support engineers and FAE deserves a special record here as well.

                      For those who prefer quick, simple and efficient device manufacturer exploring tool for your support staff - please use it directly by this link: https://flespi.io/deviceassist/. You can add the token to the link for auto-login if needed.

                      This tool can be integrated into your CRM, ERP, Support portal via iframe. Also device assistant can be integrated into your support chat or LLM as a function tool or source of information via API.

                        2 months later

                        We've enhanced codi's capabilities with protocol implementation analysis features. Now our AI assistant can:

                        • Analyze protocol implementation details
                        • Explain parameter names, values and conditions
                        • Compare protocol specifications with actual implementation in flespi

                        This allows for deeper technical discussions and more accurate troubleshooting. You can access these new capabilities by clicking the CHAT button in the top-right corner of your flespi.io account and communicating with codi.

                        11 days later

                        Codi received a per-chat memory and can activate it to access information you shared with flespi in your past conversations. This keeps its communication with the user very private and specific. You may now ask it what it knows about you ;-)

                        And we have implemented a self-training system that automatically creates anonymized long-term knowledge for the AI from all complex and resolved use cases in flespi HelpBox since 2018 and applies it when appropriate. This feature immediately pushed codi's intelligence level to new heights, making it now almost unreachable by a human in all flespi-related discussions.

                        Write a Reply...