Sound Familiar?

Ask: "How do I send emails in my PHP app?"

Naive RAG Result:

• Returns docs about the deprecated mail() function
• Mixes up SwiftMailer (abandoned) with Symfony Mailer
• Context window overloaded with irrelevant docs

"Embedding + Vector Store" – sounds simple, but it's just the beginning.

My Work in symfony/ai

The examples in this talk are based on real PRs in symfony/ai from the last 3 weeks.

Problem: Semantic Gap & Ambiguity

User writes: "how do i set up auth"

• Ambiguous: JWT? OAuth? Symfony Security?
• Implicit: Version? Bundle? Stack?
• Vector search finds "email validation" when you ask "send email"

The retriever doesn't understand the question.

Solution: Query Analysis

Techniques

• Query Rewriting – expand abbreviations, add context, fix typos
• Intent Classification – documentation vs. troubleshooting vs. comparison
• Entity Extraction – framework, version, topic

Problem: Keyword Blindness

• Vector search misses exact matches like "Mailer" or "security.yaml"
• Semantic similarity alone isn't enough
• You need both meaning and exact terms

Vector search finds similar docs – but misses the obvious ones.

Solution: Hybrid Retrieval

Vector Search + Full-Text = Better Together

Reciprocal Rank Fusion (RRF)

$\text{RRFscore}\left(d\right)=\sum _{i\in \text{Ranks}}\frac{1}{k+{\text{rank}}_i\left(d\right)}$

• d – the document being scored
• rank_i(d) – position of d in the i-th ranking (starting at 1)
• k – smoothing parameter (typically 60), prevents low ranks from dominating

Problem: Wrong Order

• Top-K isn't Top-Relevant
• Mixes SwiftMailer (abandoned) with Symfony Mailer
• Returns mail() docs alongside modern Mailer docs

We have the right candidates – but in the wrong order.

Solution: Reranking

Two-Stage Architecture

How It Works

• A Cross-Encoder takes (query, document) pairs and scores them jointly
• Unlike embeddings, it sees both texts at once → much more accurate
• Models: BAAI/bge-reranker-v2-m3 (HuggingFace) · Cohere/rerank-english-v3.0 (Cloud)

Problem: How Does RAG Reach AI Agents?

How does your RAG pipeline get into Claude/Cursor/VS Code?

Solution: MCP (Model Context Protocol)

• Your RAG pipeline becomes an MCP Tool
• AI Agents can "ask" instead of "guess"

┌─────────────┐         ┌─────────────┐
│ Claude /    │ ←──────→ │ MCP Client  │
│ Cursor      │  JSON-RPC│             │
└─────────────┘         └──────┬──────┘
                               │
                               ↓
                      ┌─────────────────┐
                      │ MCP Server      │
                      │ (your app)      │
                      └─────────────────┘

Outlook: ai-mate + Enhanced RAG

MCP server for debugging & developing PHP/Symfony apps

• Already works: ./bin/mate serve → AI agents get project context
• Coming soon: search_docs tool with Enhanced RAG (open PR)
• Goal: Your AI assistant understands your codebase and the docs behind it

Query Analysis + Hybrid Retrieval + Reranking – the full pipeline from this talk, available as MCP tool in Claude/Cursor.

Why symfony/ai?

• DI, Config, Profiler, Debug Toolbar – 100% Symfony native
• 10+ LLM providers – OpenAI, Anthropic, Mistral, Ollama, ...
• 5+ vector stores – Qdrant, Chroma, Pinecone, PostgreSQL, ...
• Query Analysis, Hybrid Retrieval, Reranking – built-in
• MCP Server – first-class integration
• ai-mate – MCP server for debugging & developing PHP/Symfony apps

Resources

• symfony/ai: github.com/symfony/ai
• ai-mate: github.com/modelflow-ai/ai-mate
• MCP: modelcontextprotocol.io
• HuggingFace: huggingface.co
• Slides: wachterjohannes.github.io/rag-beyond-hello-world
• Code: github.com/wachterjohannes/rag-beyond-hello-world