Skip to content
Sponsor

Rerank

AIfiredcore.ai.rerank

Reorder candidates by how well each one answers the query, and keep the best. Picks a rerank model, served over HTTP at the address in its endpoint setting. When no rerank service answers, the node fails with an error instead of passing the candidates on unranked.

Fired. It runs when an event or a value lands on trigger, and pulls its other inputs at that moment.

Port Type Notes
trigger event Fires the node
query text
candidates memory-set
Port Type Notes
results memory-set
trigger event
Knob Kind Default Notes
model model none stays a knob (cannot become an input)

Boltjar ships a manifest for each of these models, and the picker lists them. Picking a model adds its own settings to the node as knobs, and each of those can become an input. Models explains manifests and how to add one.

Model Id Needs Can
bge-reranker v2 m3 (HTTP) rerank/bge-v2-m3 a rerank service

A second pass after a vector search: it sends the query and the hits in candidates to a rerank service over HTTP and keeps the best ones in results. Boltjar does not run the reranker itself. Set the endpoint setting under the model picker to a service that answers POST {"query", "documents"} with a scores list or a results list of {index, score}; keep says how many results to keep. When no service answers, the node turns red and names the address.

  • Embed: Turn text into an embedding vector with the embed model you pick, the way the LLM picks its model.
  • LLM: A chat / multimodal model.
  • STT: Speech to text.
  • Tool: A tool the LLM can call.
  • Tool Args: Where a tool’s body starts.
  • TTS: Text to speech.