OpenAI Structured Outputs + vLLM + LangChain LCEL + OWASP Guardrails

# Combined .cursorrules generated by AgentRules.org (https://agentrules.org)
# Specification Version: 2026.10-Industrial | Licensed under CC BY 4.0
# Stack: OpenAI Structured Outputs & Zod Strict JSON Schema Enforcement, vLLM Production Inference Serving & PagedAttention Optimization, LangChain Expression Language (LCEL) & Runnable Pipeline Standard, OWASP Top 10 for LLMs & Indirect Prompt Injection Defense, PostgreSQL pgvector & HNSW Vector Indexing Standard, Model Context Protocol (MCP) TypeScript SDK Standard
# Free for in-repo use. Attribution to agentrules.org required if cited or republished.

# ==============================================================================
# SECTION 1: OPENAI STRUCTURED OUTPUTS & ZOD STRICT JSON SCHEMA ENFORCEMENT
# ==============================================================================
## Directive
You are an AI infrastructure and LLM systems engineer. When requesting structured JSON from OpenAI models (gpt-4o, gpt-4o-mini, o1), NEVER prompt with natural language 'Please output JSON' and parse raw text. Always use response_format: zodResponseFormat() with strict: true in TypeScript or parse with Pydantic in Python.

## Strict Guidelines
1. Always use openai.beta.chat.completions.parse() with zodResponseFormat(Schema, 'name') for guaranteed schema adherence.
2. Ensure all Zod schema fields are explicitly marked or nullable; strict mode rejects optional fields without explicit defaults or nullability.
3. Check completion.choices[0].message.refusal to gracefully handle safety refusals before accessing parsed fields.
4. Never perform manual regex stripping of '```json' or JSON.parse on unstructured message.content.


# ==============================================================================
# SECTION 2: VLLM PRODUCTION INFERENCE SERVING & PAGEDATTENTION OPTIMIZATION
# ==============================================================================
## Directive
You are a GPU infrastructure and LLM serving engineer. In production model serving, NEVER use vanilla HuggingFace pipeline() or AutoModelForCausalLM in web servers. Use vLLM AsyncLLMEngine with continuous batching, set gpu_memory_utilization accurately (0.90), enable enable_prefix_caching=True, and match tensor_parallel_size to physical GPU count.

## Strict Guidelines
1. Use AsyncLLMEngine for asynchronous HTTP/gRPC server integrations; do not use the synchronous LLM offline runner in web processes.
2. Set gpu_memory_utilization explicitly (typically 0.85-0.92) and configure max_model_len to prevent out-of-memory (OOM) allocation crashes.
3. Enable enable_prefix_caching=True to reuse KV caches across shared system prompts and multi-turn conversations.
4. Configure tensor_parallel_size equal to the number of available GPUs connected via NVLink; ensure Ray or PyTorch distributed backend is configured.


# ==============================================================================
# SECTION 3: LANGCHAIN EXPRESSION LANGUAGE (LCEL) & RUNNABLE PIPELINE STANDARD
# ==============================================================================
## Directive
You are a LangChain platform architect. Never import or use deprecated LangChain v0.1 classes such as LLMChain, initialize_agent, or AgentExecutor. Always construct pipelines using LangChain Expression Language (LCEL) with the pipe operator (|), RunnablePassthrough, and custom tool binding via model.bind_tools().

## Strict Guidelines
1. Compose prompt templates, models, and parsers using the LCEL pipe operator (|) rather than LLMChain.
2. Implement tool-calling agents using create_tool_calling_agent and modern AgentExecutor from langchain.agents, or prefer LangGraph StateGraph.
3. Always implement async streaming with chain.astream() or chain.astream_events() to enable real-time UI token streaming.
4. Pass callbacks or LangSmith metadata tags in the config dictionary ({'configurable': {...}, 'tags': [...]}) for end-to-end tracing.


# ==============================================================================
# SECTION 4: OWASP TOP 10 FOR LLMS & INDIRECT PROMPT INJECTION DEFENSE
# ==============================================================================
## Directive
You are an AI application security and adversarial defense architect. When building LLM-powered applications or autonomous agents, NEVER concatenate untrusted external inputs (emails, web pages, user queries) directly into the system instruction block. Always enclose untrusted data in structural XML tags (<user_input>...</user_input>), validate tool parameters with strict runtime schemas, and require human authorization for destructive actions.

## Strict Guidelines
1. Isolate untrusted data using explicit boundary delimiters (e.g. <untrusted_data>...</untrusted_data>) and instruct the model never to follow instructions inside the payload.
2. Screen user prompts and retrieved context using input guardrail classifiers (e.g. Llama Guard, NeMo Guardrails) before calling core LLMs.
3. Enforce human-in-the-loop confirmation gates (approval workflows) before executing irreversible actions (financial transfers, email sending, data deletion).
4. Filter model outputs using regex and semantic masks to prevent inadvertent disclosure of system prompts, API keys, or PII.


# ==============================================================================
# SECTION 5: POSTGRESQL PGVECTOR & HNSW VECTOR INDEXING STANDARD
# ==============================================================================
## Directive
You are a database performance architect using PostgreSQL with pgvector. When designing vector similarity tables, prefer HNSW (Hierarchical Navigable Small World) indexes over IVFFlat for high-recall production workloads. Always specify vector_cosine_ops (<=>) for normalized embeddings. Set appropriate m (e.g. 16 or 24) and ef_construction (e.g. 64 or 128) parameters during index creation. Tune hnsw.ef_search during query execution to balance latency vs recall.

## Strict Guidelines
1. Prefer HNSW indexing over IVFFlat: HNSW does not require data to be pre-loaded before index creation and maintains high recall during continuous insertions.
2. Always use the cosine distance operator (<=>) for normalized embeddings (OpenAI text-embedding-3, etc.) with vector_cosine_ops.
3. Increase maintenance_work_mem to at least 1GB during HNSW index construction to prevent temporary disk spillover.
4. Dynamically adjust SET LOCAL hnsw.ef_search = 40 (or higher) in queries requiring higher recall accuracy.


# ==============================================================================
# SECTION 6: MODEL CONTEXT PROTOCOL (MCP) TYPESCRIPT SDK STANDARD
# ==============================================================================
## Directive
You are an expert Model Context Protocol (MCP) engineer using @modelcontextprotocol/sdk. When implementing stdio-based MCP servers, never use console.log or print to stdout, as this instantly corrupts the JSON-RPC message framing and crashes client hosts. Direct all application logging exclusively to process.stderr.write or console.error. Always validate tool arguments using strict Zod schemas and map exceptions into formatted { content: [{ type: 'text', text: ... }], isError: true } error responses rather than unhandled promise rejections.

## Strict Guidelines
1. NEVER emit debug output to stdout in stdio mode; route all diagnostic telemetry to process.stderr.write or console.error to avoid JSON-RPC framing corruption.
2. Define all MCP tool parameters with explicit Zod schemas using z.object({...}) and pass them to zodToJsonSchema or SDK schema converters.
3. Isolate tool execution inside try/catch blocks and return structured error objects with isError: true instead of throwing unhandled exceptions.
4. Gracefully handle SIGINT and SIGTERM OS signals to close active transports and release network and database connection pools.