agent-search

agent-search supercharges your RAG flow by improving accuracy with our open-source framework, backed by the Onyx team’s work on agentic search with LangGraph: https://onyx.app/blog/agent-search-with-langgraph?ref=blog.langchain.com.

Onyx builds AI search and knowledge experiences for teams that need dependable, source-grounded answers. Agent-search distills those production learnings into a developer SDK so you can ship more reliable retrieval flows without rebuilding the orchestration stack from scratch.

Why It Is Built This Way

If you want more about the motivations behind the system, the runtime tradeoffs, and why certain design choices are opinionated, start with the live GitHub Pages essay: https://nickbohm555.github.io/agent-search/architecture.html.

TLDR

If you do not have much time to read, the basic idea is simple: take a large question, break it into subquestions, answer each one with stronger RAG techniques, optionally let a human review the subquestions, and then synthesize everything into one final answer. The flow below shows that path.

Data Flow Diagram

flowchart TD
    Q["Main question"] --> D["Decompose"]
    D --> HITL{"Subquestion HITL?"}
    HITL -->|On| REV["Review / edit subquestions"]
    HITL -->|Off| SQ["Subquestions"]
    REV --> SQ

    SQ --> S1["Subquestion 1"]
    SQ --> S2["Subquestion 2"]
    SQ --> S3["Subquestion N"]

    S1 --> QE1{"Subquery expansion?"}
    QE1 -->|On| EXP1["Expand query"]
    QE1 -->|Off| RET1["Retrieve evidence"]
    EXP1 --> RET1
    RET1 --> RR1{"Rerank?"}
    RR1 -->|On| RERANK1["Rerank results"]
    RR1 -->|Off| ANS1["Answer subquestion"]
    RERANK1 --> ANS1
    CP1["Custom prompt: subquestion answers"] -.-> ANS1
    ANS1 --> SA1["Sub-answer + citations"]

    S2 --> QE2{"Subquery expansion?"}
    QE2 -->|On| EXP2["Expand query"]
    QE2 -->|Off| RET2["Retrieve evidence"]
    EXP2 --> RET2
    RET2 --> RR2{"Rerank?"}
    RR2 -->|On| RERANK2["Rerank results"]
    RR2 -->|Off| ANS2["Answer subquestion"]
    RERANK2 --> ANS2
    CP1 -.-> ANS2
    ANS2 --> SA2["Sub-answer + citations"]

    S3 --> QE3{"Subquery expansion?"}
    QE3 -->|On| EXP3["Expand query"]
    QE3 -->|Off| RET3["Retrieve evidence"]
    EXP3 --> RET3
    RET3 --> RR3{"Rerank?"}
    RR3 -->|On| RERANK3["Rerank results"]
    RR3 -->|Off| ANS3["Answer subquestion"]
    RERANK3 --> ANS3
    CP1 -.-> ANS3
    ANS3 --> SA3["Sub-answer + citations"]

    SA1 --> SYN["Final synthesis"]
    SA2 --> SYN
    SA3 --> SYN
    CP2["Custom prompt: synthesis"] -.-> SYN
    SYN --> OUT["Final answer"]

Loading

SDK Quick Reference (PyPI)

For the full, canonical SDK docs, see https://pypi.org/project/agent-search-sdk/.

The PyPI package is an in-process Python SDK for agent-search. It is intentionally narrow: consumers should call advanced_rag(...) and treat that as the supported entrypoint. The SDK always requires both:

• A chat model (for example langchain_openai.ChatOpenAI)
• A vector store that implements similarity_search(query, k, filter=None)

It does not auto-build these dependencies for you.

Install (PyPI)

python3.11 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install agent-search-sdk
python -c "import agent_search; print(agent_search.__file__)"

Quick Start

from langchain_openai import ChatOpenAI
from agent_search import advanced_rag
from agent_search.vectorstore.langchain_adapter import LangChainVectorStoreAdapter

vector_store = LangChainVectorStoreAdapter(your_langchain_vector_store)
model = ChatOpenAI(model="gpt-4.1-mini", temperature=0.0)

response = advanced_rag(
    "What is pgvector?",
    vector_store=vector_store,
    model=model,
)
print(response.output)

Included Features

• Multi-step agentic retrieval: Breaks a large question into subquestions, works them in parallel, and synthesizes a final answer from grounded evidence.
• Subquestion HITL review: Adds one high-leverage human review point so operators can inspect or edit subquestions before the expensive work continues.
• Advanced retrieval controls: Supports optional query expansion and reranking so you can trade off recall, precision, and cost per run.
• Custom prompts: Lets you override subanswer and synthesis instructions so you can tune output behavior without replacing the runtime’s evidence wiring.
• Resumable runtime: Supports checkpointed pause-and-resume flows for HITL runs so work can continue without restarting the graph.