rag-code-gen

RAG-grounded code generation

attune-rag ships with attune-ai core (pip install attune-ai). Grounds code generation in the attune-help template corpus so outputs cite real APIs, workflow names, and patterns instead of hallucinating them. Every output ends with a ## Sources block of clickable citations.

Scoping

Before running, ask:

1. What are you trying to produce? Code, config, explanation, or a mix?
2. Any specific attune surface? e.g. "the security-audit workflow", "MCP tool pattern", "BaseWorkflow subclass". Helps retrieval hit the right concept file.
3. Depth? Default is standard. quick saves time and budget for simple asks; deep is for complex multi-file or architectural questions.

Running

attune workflow run rag-code-gen \
  --input '{"query": "<the request>"}'

Optional inputs:

• k — max grounding docs to retrieve (default 3)
• depth — quick / standard / deep
• feedback — pass good or bad AFTER inspecting the output to record a verdict against every cited template
• model — override the generator model if you need to

Output shape

WorkflowResult.final_output is a string with two parts:

1. The generated code / explanation
2. A ## Sources section listing each cited attune-help template with its category, retrieval score, and a clickable link to the source on GitHub

WorkflowResult.metadata carries the full citation dict (query, retriever_name, retrieved_at, hits[]) plus fallback_used and confidence so callers can make routing decisions.

When RAG doesn't help

If the retriever can't find relevant templates, the workflow still runs — it falls back to an unaugmented prompt that explicitly tells the model there is NO grounding context so it should say "I don't know about attune X" rather than invent. metadata.fallback_used is True in that case.

Alternative: MCP tool only (no LLM call)

If you want just retrieval without generation, use the rag_knowledge_query MCP tool directly. It returns hits + an augmented prompt string without calling an LLM itself — handy for routing decisions, agent planning, or feeding another model.