Skip to main content
chat() takes config objects that control how it handles conversation flow and low-quality retrieval. Each targets a specific failure mode.

​
Conversation management

​
History compaction

Long conversations overflow the LLM’s context window and cost more on every turn. Compaction summarizes older messages into a short recap and keeps recent turns verbatim. 
from ragrails import HistoryCompactionConfig

rag.chat(..., history_compaction=HistoryCompactionConfig(
    enabled=True,
    history_limit=15,   # summarize once history passes this many turns
    keep_recent=5,      # keep the last 5 turns verbatim
))
Compaction is also a cost lever: fewer tokens per turn means lower spend on long chats. See cost optimization.

​
Intent routing

Not every message needs a database lookup. “Thanks!” or “who are you?” should get a direct reply. Intent routing detects these and skips retrieval, which is faster, cheaper, and avoids irrelevant context. 
from ragrails import IntentRoutingConfig

rag.chat(..., intent_routing=IntentRoutingConfig(enabled=True))  # on by default
result.intent is "rag" (retrieval ran) or "direct" (answered without retrieval).

​
Answer quality

Sometimes your index simply doesn’t contain the answer. Rather than hallucinate, Ragrails scores the retrieved context and acts on low confidence. 
from ragrails import ChatRetrievalQualityConfig

rag.chat(..., retrieval_quality=ChatRetrievalQualityConfig(
    min_retrieval_score=0.35,
    min_rerank_score=0.50,
    low_confidence_mode="answer_with_caution",
    max_context_chunks=None,
))
low_confidence_modeBehaviour when context is weak
answer_with_cautionAnswers, but flags the uncertainty (default)
ask_clarifying_questionAsks the user to refine the question instead of answering
refuse_grounded_answerDeclines to answer rather than risk a wrong one
Every response carries result.answer_confidence and result.retrieval_quality. Surface them in your UI to flag shaky answers. Reference: SDK chat.