The Intent Blade: SLM Orchestration

The Intent Blade is the conversational intelligence layer of ANDARTIS. Rather than using fixed rule parsers, it runs a Small Language Model (SLM) locally on Apple Silicon to understand complex prompts, select tools, and format raw database records.

---

Technical Stack

• ML Library: Apple mlx-lm running in the persistent Python daemon.
• Model Target: High-performance, quantized 7B-class models (e.g., Mistral-7B-Instruct-v0.3-4bit) or ultra-lightweight 3B models.
• VRAM Persistence: The weights remain loaded in Apple Silicon's Unified Memory, ensuring near-instant response times for subsequent queries.
• Storage Path: Configured globally inside storage/models/global_slm/.

---

Execution Phases

flowchart TD
    Query[User Query] --> Intent[Intent Parsing Phase]
    Intent --> |JSON Schema Plan| Tools{Tool Orchestrator}
    Tools --> |SQL Query| DB[(SQLite DB)]
    Tools --> |Neural Search| Vector[Semantic Navigator]
    DB --> Output[Raw JSON Data]
    Vector --> Output
    Output --> Synthesis[Grounded Synthesis Phase]
    Synthesis --> Response[Conversational prose]

1. Intent Parsing

When you submit a prompt, the SLM interprets your intent and compiles a structured JSON execution plan containing specific tools and SQL filters.

Supported Actions

• hybrid_search: Combines a strict metadata SQL filter with semantic neural ranking. Recommended for scaling across 10,000+ files.
• entity_analytic: Used to run high-speed quantitative counting or database statistics.
• semantic_search: Used to surface abstract concepts across different files.
• aggregate_query: Processes map-reduce style extractions over large datasets.
• clarify: Prompted when a user query is ambiguous, generating a professional clarifying question.

---

Grounded Synthesis (The Model Weaver)

To process massive contexts without exceeding local RAM limits, ANDARTIS uses the Model Weaver pipeline:

1. Information Filtering (Micro): The Semantic Navigator or Entity Analytic executes the initial query. For a dataset of 10,000 documents, it filters the set down to approximately 50 candidates using metadata, then ranks the top 5 semantically.
2. Context Compilation: The PHP Orchestrator retrieves the text of the top 5 chunks and structures them into a raw "Ground Truth Context".
3. The Synthesis Pass: The context is fed to the Senior Analyst (SLM) with the synthesize action. The model writes the final response using only the provided facts, ensuring zero hallucinations.

---

Sample Interaction Flow

Below is a trace of the system processing the query: “Show me patient summaries from March”

1. Input: User sends the query via the Vue UI.
2. Parsing: The Intent Blade parses the query and returns a structured JSON payload:

   {
     "action": "aggregate_query",
     "query": "SELECT summary, date FROM clinical_records WHERE date LIKE '2026-03-%'",
     "parameters": {}
   }

3. Retrieval: The Laravel core runs the SQL query against the SQLite database in milliseconds, retrieving matching rows.
4. Packaging: The core packages the raw results and forwards them to the Python synthesis pipeline.
5. Synthesis: The SLM receives the JSON array and compiles it into a bulleted summary.
6. Output: The user sees a formatted response in the desktop window.