vcase-auto-context 0.1.1

Version-Constrained API Specification Engine (VCASE)

Version-Constrained API Specification Engine

A deterministic, version-aware knowledge streaming infrastructure for Large Language Models (LLMs).

Why does this exist?

When LLMs generate code, they frequently rely on outdated training data or hallucinate API signatures for newer libraries. Traditional Retrieval-Augmented Generation (RAG) is often static and misses critical version-specific nuances or architectural constraints.

This engine replaces static RAG with a five-layer defense system. It dynamically extracts exact project dependencies, resolves them to specific versions, and streams deterministic API contracts and interaction rules directly into the LLM's system prompt.

Core Features

• Zero-Hallucination API Contracts: Injects exact function signatures based on the specific version of the library you are using.
• Architectural Guardrails: Enforces complex interaction patterns (e.g., "Temporal workflows cannot be executed directly inside FastAPI route handlers").
• Language Agnostic Architecture: Built on a provider interface allowing expansion beyond Python to Node, Rust, Go, etc.

Resolution Architecture

The engine uses a strict Tiered Fallback Strategy to guarantee dependency accuracy before building the specification:

1. Tier 1: Lock Files (Highest Priority)
   • Parses deterministic manifests (e.g., poetry.lock) to extract exact, pinned versions.
2. Tier 2: Declaration Files
   • Parses abstract declarations (e.g., pyproject.toml, requirements.txt).
   • Connects to ecosystem registries (like PyPI) to resolve loose constraints (like ^1.2.0) to the latest compatible version.
3. Tier 3: Containerized Introspection (Future)
   • Sandboxed subprocess execution (pip freeze, npm ls) for deeply nested or hidden dependencies.

Component Flow

1. TechnologyDetector: Scans the repository or user prompt against a curated list of ~180 production-grade technologies to identify the stack.
2. DependencyResolver: Orchestrates the tiered fallback logic across all available Ecosystem Providers.
3. EcosystemProvider (e.g., PythonProvider): Parses the specific manifest syntax of the language.
4. PypiClient: Makes async, defensive network requests to fetch current metadata and documentation sources.
5. SpecBuilder: Assembles the final, non-negotiable System Prompt block, formatting parameters, return types, and deprecation warnings.

Usage / Integration

Currently, the engine provides internal APIs to assemble the System Prompt constraints.

# Example Internal Flow (Pseudocode)
from vcase.service.resolver import DependencyResolver
from vcase.service.spec_builder import SpecBuilder

# 1. Resolve exact versions used in the project
resolver = DependencyResolver(providers=[PythonProvider(...)])
resolved_deps = await resolver.resolve(Path("./my_project"))

# 2. Fetch the metadata/signatures internally 
# ...

# 3. Build the strict constraints for the LLM
builder = SpecBuilder()
system_instruction = builder.build_system_prompt(api_indexes, patterns)

print(system_instruction) 
# Outputs: 
# ## litellm 1.40.0
# - completion(model: str, messages: list) -> ModelResponse
# ⚠ PATTERN: fastapi + temporalio
# Rule [NO_WORKFLOW_IN_HANDLER]: Workflows must run in workers.

Contributing

When adding a new Ecosystem Provider (e.g., NodeProvider), you must adhere to the asynchronous contract defined in vcase/providers/base.py. All registry interactions must be defensive and non-blocking.