ZeroPrompt.Core 1.3.0

🎯 ZeroPrompt: Sovereign Pure C# Prompt Templating & Grammar-Constrained Engine

ZeroPrompt is an industrial-grade, zero-allocation prompt templating, Pushdown Automaton (PDA) JSON Schema state machine, Grammar-Constrained Logit Masker, and dynamic few-shot exemplar selector written in 100% pure C#. It operates at Tier 5 (Presentation & Orchestration) of the ZeroPlatform ecosystem.

⚡ Key Capabilities

• Zero-Allocation Prompt Templating (PromptTemplate):
  • Blazing-fast variable interpolation syntax {{variable_name}} without heavyweight regex or runtime reflection.
  • Span-based rendering directly into StringBuilder buffers.
  • Parameter validation and missing variable tracking.
• Pushdown Automaton (PDA) JSON Grammar Engine (JsonGrammarState):
  • Microsecond deterministic stack-based validation of LLM generation streams.
  • Tracks nested objects ({), arrays ([), string literals, escape sequences, colons, and commas.
• Schema-Constrained JSON Pushdown Automaton (SchemaConstrainedJsonGrammar):
  • 64-bit Bitmask Property Validation: Maps property indices to 64-bit unsigned integers (ulong), enforcing property whitelists, required property completeness, and strict data type constraints with single-cycle bitwise operations.
  • Zero Heap Allocations: Replaces heap-allocated bracket stacks with a fixed inline buffer (char[16]), reducing token candidate testing latency to under $0.5,\mu\text{s}$.
• Strict Tool Calling Finite Automaton (ToolCallGrammarState):
  • Governs the <tool_call>{"tool": "<name>", "parameters": { ... }}</tool_call> syntax.
  • Zero Tool Name Hallucination: Only allows the generation of tool names explicitly registered in the agent's tool catalog. Automatically switches parameter schemas upon tool name resolution.
• Context-Aware Grammar Logit Processor (ToolCallGrammarLogitProcessor):
  • Integrates directly with ZeroLlm via ContextAwareLogitProcessor. Dynamically sets invalid syntax/schema token logits to $-\infty$ (-1e9f), mathematically guaranteeing 100% syntactically valid JSON tool calls.
• Dynamic Few-Shot Exemplar Selector (FewShotSelector):
  • Backed by ZeroVector.Core hardware-accelerated SIMD vector similarity (VectorMetrics.CosineSimilarity).
  • Dynamically selects the top-$K$ most semantically relevant few-shot demonstrations for prompts.
• Transformer KV-Cache Layout Optimizer (PromptLayoutOptimizer):
  • Partitions prompt components strictly into cacheable static prefixes (StaticSystem, ToolDefinitions, FewShot, ContextRAG) and dynamic request-specific payloads (History, DynamicSuffix), cutting TTFT by up to 70%.
• Zero External Dependencies & Multi-Targeting:
  • Compatible with .NET 8.0+, .NET Framework 4.6.2+, and .NET Standard 2.0.

🚀 Quick Start

1. Template Rendering

using ZeroPrompt.Core.Templates;

var template = new PromptTemplate("You are a SCADA diagnostic engineer for line {{line_id}}. Analyze status: {{status}}.");
var context = new PromptContext()
    .Set("line_id", "LINE-04")
    .Set("status", "OVERHEAT_ALARM");

string prompt = template.Render(context);
Console.WriteLine(prompt);

2. Schema-Constrained JSON Grammar

using ZeroPrompt.Core.Grammar;

var schema = new JsonSchemaConstraint("so_query")
    .AddProperty("order_id", SchemaPropertyType.String, required: true)
    .AddProperty("limit", SchemaPropertyType.Number, required: false);

var grammar = new SchemaConstrainedJsonGrammar(schema);

// Validate character transitions with zero heap allocation
bool valid = grammar.CanAcceptNext("{\"order_id\": \"SO-2026-001\"}".AsSpan());
Console.WriteLine($"Valid: {valid}");

3. Tool Call Grammar Logit Masking

using ZeroPrompt.Core.Grammar;

var schemas = new Dictionary<string, JsonSchemaConstraint>
{
    ["mds_db_so_query"] = new JsonSchemaConstraint("mds_db_so_query")
        .AddProperty("order_id", SchemaPropertyType.String, required: true)
};

var processor = new ToolCallGrammarLogitProcessor(tokenizer, schemas);

// Applied directly during LLM autoregressive token sampling
processor.Process(logitsSpan, pastTokensSpan);

🛡️ Architecture & Tier Compliance

flowchart TD
    PromptTemplate["ZeroPrompt.Core.Templates\n(PromptTemplate, PromptContext)"]
    JsonGrammar["ZeroPrompt.Core.Grammar\n(SchemaConstrainedJsonGrammar, ToolCallGrammarState)"]
    LogitMasker["ZeroPrompt.Core.Grammar\n(ToolCallGrammarLogitProcessor)"]
    FewShot["ZeroPrompt.Core.FewShot\n(FewShotSelector, Exemplar)"]
    
    FewShot --> VectorMetrics["ZeroVector.Core.Metrics\n(SIMD CosineSimilarity)"]
    LogitMasker --> ZeroLlm["ZeroLlm.Core.Sampling\n(ContextAwareLogitProcessor)"]
    JsonGrammar --> ZeroAgent["ZeroAgent.Core\n(AgentToolRegistry)"]

📄 License

Architected and developed by Phong Võ (kzxl) for the ZeroUniverse / ZeroPlatform ecosystem. Released under the MIT License.