LangChain

By reciprocal rank fusion, not by score arithmetic. EnsembleRetriever runs each wrapped retriever, converts every result to its rank position, sums weight divided by (c + rank) across retrievers, and returns the deduplicated candidates ordered by that fused score.

MultiQueryRetriever asks an LLM to rewrite the question into several phrasings, runs the base retriever on each, and returns the deduplicated union. It buys recall against vocabulary mismatch; it costs one extra LLM call of latency and a larger, unranked document set.

Use a fixed pipeline whenever the sequence of steps is known before the request arrives. An agent spends model calls deciding what you already know, adds latency variance and non-determinism, and makes evaluation and incident response markedly harder.

When the workflow needs cycles, durable state between turns, mid-run human approval, or per-step recovery. LCEL composes a directed, stateless pipeline; those requirements need a graph runtime with checkpointing, or plain imperative code.

Treat the get_session_history factory as the seam and pick the backend from the requirements behind it: latency and TTL push toward a cache, durability and audit toward a database. Key by authenticated user plus conversation, bound each read, and set retention and deletion policy alongside.

Construct models from configuration with init_chat_model rather than importing provider classes at call sites, keep a small set of named tiers per task, and gate any switch behind evaluation — because capability differences leak through the shared interface even when the code compiles.