Skip to main content
Chat is stateless. Pass history in on each turn and persist the returned result.history in your application. 
pip install "ragrails[voyage,openai,qdrant]"

from ragrails import RagRails

rag = RagRails()
llm = rag.llm(provider="openai", model="gpt-4o-mini")
embedder = rag.embedder(provider="voyage", model="voyage-3", input_type="query")

history = []

result = rag.chat(
    "How do I authenticate?",
    llm=llm,
    embedder=embedder,
    vector_db="qdrant",
    collection="docs",
    url="http://localhost:6333",
    history=history,
)

print(result.answer)
history = result.history  # pass to the next turn

​
Result fields

FieldDescription
result.answerLLM answer string
result.sourcesSource chunk dicts used for the answer
result.historyUpdated history list; pass to the next turn
result.intent"rag" (retrieved context used) or "direct" (LLM answered directly)
result.answer_confidenceConfidence assessment dict
result.retrieval_qualityRetrieval quality assessment dict
result.compactedTrue if history was summarised this turn
result.errorsList of error dicts

​
Config objects

​
QueryRewriteConfig

Rewrites follow-up questions into self-contained queries before retrieval. 
from ragrails import QueryRewriteConfig

result = rag.chat(
    "What about the second step?",
    ...,
    query_rewrite=QueryRewriteConfig(
        enabled=True,
        session_context="User is asking about the onboarding flow.",
    ),
)
FieldDefaultDescription
enabledFalseEnable query rewriting
session_context""Context hint for the rewriter
llmNoneSeparate LLM for rewriting (falls back to the chat LLM)

​
HistoryCompactionConfig

Automatically summarises old history turns to stay within context limits. 
from ragrails import HistoryCompactionConfig

result = rag.chat(
    ...,
    history_compaction=HistoryCompactionConfig(enabled=True, history_limit=15, keep_recent=5),
)
FieldDefaultDescription
enabledTrueEnable automatic history summarisation
history_limit15Summarise when history exceeds this many turns
keep_recent5Recent turns to keep verbatim after summarisation

​
IntentRoutingConfig

Routes small-talk and direct questions to the LLM without retrieval. 
from ragrails import IntentRoutingConfig

result = rag.chat(..., intent_routing=IntentRoutingConfig(enabled=True))
FieldDefaultDescription
enabledTrueRoute non-RAG queries directly to the LLM

​
ChatRetrievalQualityConfig

Controls confidence thresholds for retrieved context. 
from ragrails import ChatRetrievalQualityConfig

result = rag.chat(
    ...,
    retrieval_quality=ChatRetrievalQualityConfig(
        min_retrieval_score=0.35,
        min_rerank_score=0.50,
        low_confidence_mode="answer_with_caution",
    ),
)
FieldDefaultDescription
min_retrieval_score0.35Below this, retrieval is considered low quality
min_rerank_score0.50Below this, reranked results are low quality
low_confidence_mode"answer_with_caution""answer_with_caution", "ask_clarifying_question", or "refuse_grounded_answer"
max_context_chunksNoneCap chunks sent to the LLM

​
llm() parameters

ParameterDefaultDescription
provider"""openai", "anthropic", or "google" (inferred from the model when omitted)
modelrequiredModel name, e.g. "gpt-4o-mini", "claude-sonnet-4-6", "gemini-2.5-flash"
max_tokens1024Max output tokens

​
chat() parameters

ParameterDefaultDescription
queryrequiredUser message
llmrequiredLLM object from rag.llm()
embedderrequiredEmbedder object from rag.embedder() with input_type="query"
vector_db"qdrant"Vector DB provider
collectionNoneCollection name
urlNoneVector DB URL
rerankerNoneReranker object from rag.reranker()
historyNonePrevious turns list
history_compactionNoneHistoryCompactionConfig
query_rewriteNoneQueryRewriteConfig
intent_routingNoneIntentRoutingConfig
retrieval_qualityNoneChatRetrievalQualityConfig
persona""System persona injected into the prompt
optionsNoneProvider-specific options dict