Evolving into a Neuro-Symbolic Agent: neo-pkg-llm-chat 3.0.0 Released
The last weakness of 2.x was code written directly by AI.
Version 2.0.0 greatly reduced failures by separating features according to intent and automatically correcting invalid input.
But one class of problem remained: chart code written directly by AI.
Creating a chart requires query code behind the scenes, and that code contains many pitfalls.
The AI repeatedly made the same mistakes in handling time values and applying aggregation and interval rules.
Lightweight local models, in particular, often produced charts that failed to run or charts with valid syntax but no data. In worse cases, the AI invented code, presented it as the solution, and gave the user code that was actually incorrect.
The biggest change in 3.0.0—the neuro-symbolic agent—addresses this problem directly.
Neuro-symbolic agent
New forecasting feature (Beta)
Automatic blocking of incorrect answers
Stronger security
More accurate document search
Per-user favorites
Stability and usability improvements
1. Neuro-symbolic agent: AI captures intent; the symbolic engine writes the code
We no longer ask AI to write chart code directly.
Instead, AI describes the intent—what to view, over what period, and how to group it—while a rule-based symbolic engine generates the code that actually runs.
The key is to separate understanding natural language and extracting intent (the neural part) from converting that intent into code that avoids known pitfalls (the symbolic part). AI, which can be inaccurate, handles what it does well: understanding intent. The symbolic engine takes responsibility for code generation, where correctness can be explicitly defined.
The generated code is also executed once for validation before it is shown to the user.
This applies both when rendering a chart and when an answer provides code to explain how to query the data.
As a result, within the scope supported by the engine, charts that cannot run cannot reach the user. The symbolic engine handles the pitfalls of code generation, while AI is responsible for intent.
2. Forecasting: New Beta feature
*This is a Beta feature. It currently focuses on linear trend forecasting, and its accuracy is still being validated across a wider range of data. We recommend using it as a guide to trends, rather than as the sole basis for operational decisions.
The agent can now answer questions about what may happen next.
It predicts future values from recent trends and plots confidence intervals to help users assess uncertainty.
A straight trend line is the default. Simple models are prioritized to avoid excessive extrapolation. Curved or cyclical models are used only when the data shows clear acceleration, deceleration, or recurring patterns, and those decisions are currently made conservatively.
Forecasts are displayed with confidence intervals.
When added to a dashboard, a forecast is recalculated using the latest data each time the chart is opened, so it stays current.
If there are too many tags, the agent asks what to forecast instead of plotting them all at once.
3. Blocking incorrect answers before they reach the user
While the neuro-symbolic agent and forecasting features focus on producing correct results,
this layer prevents incorrect results from slipping through. It automatically filters the AI mistakes most frequently encountered during operation.
If the AI tries to include invented code in an answer, the system blocks it and requires regeneration.
If the AI finishes an analysis but forgets to create the dashboard, the system completes that step.
Unverified or incorrect charts are not included in answers.
The principle is simple: results shown to users must have been validated by the system.
4. Security: Consistent blocking of dangerous commands
Attempts to modify or delete data, control the server, or execute system commands are blocked consistently at a central point.
The same rules apply regardless of the AI model or the path through which a request arrives.
Queries, analysis, and normal data collection continue to work, while dangerous operations are refused with guidance to perform them directly in the console.
5. More accurate document search
Previously, manual searches sometimes read only the beginning of a document, leading the AI to incorrectly state that a feature described later was unsupported. Version 3.0.0 improves the ability to locate and read the relevant sections, reducing hallucinations that deny the existence of supported features.
6. Per-user favorites
Each user can save frequently used items separately.
Building on the separation of per-user settings in 2.0.0, this expands personalization.
7. Stability and usability improvements
We have fixed bugs found in everyday use, including repeated answers, dashboards that would not open under certain conditions, and simple query results shown as charts instead of tables.
We have also added examples based on the questions most frequently asked in our regular training sessions. Users can select examples by topic in LLM Chat and run queries immediately.
This makes it easier for first-time users to find and use the features they need.
At a glance: 2.5.0 → 3.0.0
Feature | 2.5.0 | 3.0.0 |
Chart code | Written directly by AI | AI captures intent; the symbolic engine generates validated code |
Forecasting | None | Trend forecast charts (mainly linear trends with confidence intervals) — Beta |
Incorrect answers | Could reach the user | Only validated results are shown |
Security | Handled in multiple places | Consistently blocked at one central point |
Document search | Read only the beginning, leading to incorrect answers | Finds and reads the relevant sections |
Per-user features | Separate settings | Separate settings and favorites |
Closing thoughts
If 2.0.0 refined the architecture, 3.0.0 builds on that architecture to reduce AI guesswork and add new capabilities.
Our guiding principle is simple: do not leave tasks with clearly defined correct answers and many pitfalls to AI.
AI expresses what the user wants to see. The symbolic engine is responsible for turning that intent into code that avoids known pitfalls, then executing and validating it.
On top of that, we have added forecasting capabilities and improved operational trust through security, document accuracy, and user isolation.
We will continue improving the agent so that it can be relied on in real work.
![[Edge Computing Series #3] The Architecture of Micro Data Lakes](https://static.wixstatic.com/media/adc7e5_c3774374c53f40f1a91e40fbc60eaee1~mv2.png/v1/fill/w_980,h_550,al_c,q_90,usm_0.66_1.00_0.01,enc_avif,quality_auto/adc7e5_c3774374c53f40f1a91e40fbc60eaee1~mv2.png)
![[Smart Factory Basics #1] ISA-95: An International Framework](https://static.wixstatic.com/media/adc7e5_5764c89cc04c4e519e58381466cc64fa~mv2.jpg/v1/fill/w_501,h_373,al_c,q_80,enc_avif,quality_auto/adc7e5_5764c89cc04c4e519e58381466cc64fa~mv2.jpg)
