rasa-setting-up-enterprise-search

Setting Up Enterprise Search

Enterprise Search lets a Rasa assistant answer informational questions by retrieving relevant documents from a knowledge base and generating (or extracting) answers. It is powered by the EnterpriseSearchPolicy and triggered via the built-in pattern_search pattern.

Workflow

1. Check the existing project — does it already have EnterpriseSearchPolicy configured? If not, add EnterpriseSearchPolicy to policies in config.yml with the desired vector store type and options (see "Enterprise Search Policy").
2. Switch the command generator in config.yml to SearchReadyLLMCommandGenerator (see "Command generator").
3. Add vector store connection details to endpoints.yml if using Milvus or Qdrant (see "Connecting vector stores").
4. Add model groups for the LLM and embeddings used by the policy (see rasa-configuring-model-groups skill).
5. Override pattern_search in a flows file to trigger action_trigger_search (see "Triggering Enterprise Search").
6. Place documents in the knowledge base — ./docs for Faiss, or ingest into your vector database for Milvus/Qdrant.
7. Validate and train.

Command generator

Enterprise Search requires SearchReadyLLMCommandGenerator in the pipeline. This is the search-aware variant of CompactLLMCommandGenerator — it produces SearchAndReply commands when the LLM within the Command Generator detects an informational question.

If the project currently uses CompactLLMCommandGenerator, replace it.

# config.yml
pipeline:
  - name: SearchReadyLLMCommandGenerator
    llm:
      model_group: command_generator_llm    # defined in endpoints.yml

Enterprise Search Policy

Add EnterpriseSearchPolicy to policies in config.yml. The policy supports two modes:

• Generative (default) — uses an LLM to produce a context-aware answer from retrieved documents. When check_relevancy is enabled and the answer is not relevant, the policy triggers pattern_cannot_handle.
• Extractive — set use_generative_llm: false to return a pre-authored answer directly with no LLM generation. Documents must be ingested in Q&A format (see example below). Use vector_store.threshold so only high-confidence matches are returned.