perf: optimize Docker build time from >10min to ~2min

- Remove sentence-transformers dependency (saves ~3GB PyTorch/CUDA)
- Make CrossEncoder import optional with graceful fallback
- Optimize Dockerfile for layer caching (incremental builds ~3s)
- Change PostgREST port from 3000 to 3001 (avoid Next.js conflict)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

compose.yaml

@@ -51,7 +51,7 @@ services:
       PORT: 8051
       # Embedding config - point to remote LiteLLM
       EMBEDDING_API_BASE: https://llm.ai.pivoine.art/v1
-      EMBEDDING_MODEL: bge-large-en-v1.5
+      EMBEDDING_MODEL: hosted_vllm/BAAI/bge-large-en-v1.5
       EMBEDDING_DIMENSION: "1024"
       OPENAI_API_KEY: ${LITELLM_API_KEY}
       # Supabase-compatible config (pointing to PostgREST)

config.toml

@@ -1,6 +1,6 @@
 model_provider = "litellm"
 # model = "anthropic/claude-sonnet-4-5-20250929"
-model = "hosted_vllm/openai/qwen-2.5-7b"
+model = "hosted_vllm/meta-llama/Llama-3.1-8B-Instruct"
 
 [projects."/home/valknar"]
 trust_level = "trusted"

servers/mcp-crawl4ai-rag/Dockerfile

@@ -1,21 +1,26 @@
 FROM python:3.12-slim
 
 ARG PORT=8051
-
 WORKDIR /app
 
-# Install uv
+# Install uv first (cached layer)
 RUN pip install uv
 
-# Copy the MCP server files
+# Copy only dependency files first (cached layer)
+COPY pyproject.toml uv.lock* ./
+
+# Install dependencies (cached unless pyproject.toml changes)
+RUN uv pip install --system .
+
+# Run crawl4ai-setup for Playwright (cached layer)
+RUN crawl4ai-setup
+
+# Copy source code last (only this invalidates on code changes)
 COPY . .
 
-# Install packages directly to the system (no virtual environment)
+# Re-install in editable mode (fast, deps already installed)
-# Combining commands to reduce Docker layers
+RUN uv pip install --system -e . --no-deps
-RUN uv pip install --system -e . && \
-    crawl4ai-setup
 
 EXPOSE ${PORT}
 
-# Command to run the MCP server
 CMD ["python", "src/crawl4ai_mcp.py"]

servers/mcp-crawl4ai-rag/pyproject.toml

@@ -10,6 +10,5 @@ dependencies = [
     "supabase==2.15.1",
     "openai==1.71.0",
     "dotenv==0.9.9",
-    "sentence-transformers>=4.1.0",
     "neo4j>=5.28.1",
 ]

servers/mcp-crawl4ai-rag/src/crawl4ai_mcp.py

@@ -6,8 +6,15 @@ the appropriate crawl method based on URL type (sitemap, txt file, or regular we
 Also includes AI hallucination detection and repository parsing tools using Neo4j knowledge graphs.
 """
 from mcp.server.fastmcp import FastMCP, Context
-from sentence_transformers import CrossEncoder
 from contextlib import asynccontextmanager
+
+# Optional import for reranking functionality
+try:
+    from sentence_transformers import CrossEncoder
+    RERANKING_AVAILABLE = True
+except ImportError:
+    CrossEncoder = None
+    RERANKING_AVAILABLE = False
 from collections.abc import AsyncIterator
 from dataclasses import dataclass
 from typing import List, Dict, Any, Optional
@@ -146,14 +153,16 @@ async def crawl4ai_lifespan(server: FastMCP) -> AsyncIterator[Crawl4AIContext]:
     # Initialize Supabase client
     supabase_client = get_supabase_client()
     
-    # Initialize cross-encoder model for reranking if enabled
+    # Initialize cross-encoder model for reranking if enabled and available
     reranking_model = None
-    if os.getenv("USE_RERANKING", "false") == "true":
+    if os.getenv("USE_RERANKING", "false") == "true" and RERANKING_AVAILABLE:
         try:
             reranking_model = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
         except Exception as e:
             print(f"Failed to load reranking model: {e}")
             reranking_model = None
+    elif os.getenv("USE_RERANKING", "false") == "true" and not RERANKING_AVAILABLE:
+        print("Reranking requested but sentence-transformers not installed - skipping")
     
     # Initialize Neo4j components if configured and enabled
     knowledge_validator = None