Deep Searcher: An Open‑Source Agentic RAG Framework for Enterprise‑Level Search and Knowledge Retrieval

With the rapid evolution of AI technologies, search applications have become the first consensus layer of AI usage. OpenAI, Microsoft Bing Copilot, Perplexity AI, and domestic products like Doubao and Kimi exemplify this trend. However, users now expect large language models to handle enterprise‑level scenarios, not just simple keyword retrieval.

OpenAI recently released DeepResearch, a tool that can generate comprehensive analysis reports from a single query such as “What is a reasonable market cap for Tesla?”. Inspired by this, Zilliz launched the open‑source Deep Searcher project, which has quickly amassed over 3,100 GitHub stars.

Deep Searcher embodies an Agentic RAG architecture, where a large model actively plans, iterates, and refines its retrieval process. Compared with traditional RAG, Agentic RAG offers three major advantages: (1) passive response becomes proactive response, (2) single‑step keyword search evolves into multi‑round dynamic retrieval with self‑correction, and (3) simple fact‑questioning expands to complex reasoning and report generation.

The system consists of two main modules. The data‑ingestion module uses Milvus vector database to import private knowledge bases, making Deep Searcher suitable for enterprises with exclusive data. The online inference module implements various agent strategies and the RAG pipeline, enabling dynamic looped iterations: after each retrieval, a reflection step evaluates whether the knowledge is sufficient; if not, a new query is generated and the cycle repeats until a final report is produced.

Example workflow: a query about the evolution of The Simpsons is broken into sub‑queries, each sent to the vector store, with the LLM iteratively refining the sub‑questions and aggregating results into a detailed report. The article also shows a complex logical query about film directors, illustrating how Deep Searcher decomposes the problem, retrieves director information, compares birth dates, and arrives at the correct answer, whereas a vanilla RAG may hallucinate.

Quantitative evaluations on the 2WikiMultiHopQA dataset demonstrate that increasing the maximum number of feedback iterations improves recall, though marginal gains diminish after a few rounds. Token consumption grows linearly with iteration count, costing roughly $0.0036 per 50‑query batch using gpt‑4o‑mini pricing. Model comparisons reveal that inference‑oriented models (e.g., Claude 3.7 Sonnet) outperform non‑inference models, but weaker models may fail to complete the agentic workflow.

When contrasted with Graph RAG, Agentic RAG (Deep Searcher) shows greater flexibility: Graph RAG requires relationship extraction during data ingestion and performs fixed graph traversals at query time, while Agentic RAG postpones heavy reasoning to the query phase, allowing dynamic routing and multi‑step problem solving.

Finally, the article reflects on practical lessons: inference models are powerful but not universally optimal; token budgets impose diminishing returns; traditional RAG is limited to shallow retrieval; and Agentic RAG, powered by cost‑effective LLMs and efficient vector search, is poised to become the dominant paradigm for deep content generation and enterprise knowledge retrieval.

Code examples from the article:

Break down the original query into new sub queries: [
  'How has the cultural impact and societal relevance of The Simpsons evolved from its debut to the present?',
  'What changes in character development, humor, and storytelling styles have occurred across different seasons of The Simpsons?',
  'How has the animation style and production technology of The Simpsons changed over time?',
  'How have audience demographics, reception, and ratings of The Simpsons shifted throughout its run?'
]

Next‑iteration sub‑queries:

New search queries for next iteration: [
  "How have changes in The Simpsons' voice cast and production team influenced the show's evolution over different seasons?",
  "What role has The Simpsons' satire and social commentary played in its adaptation to contemporary issues across decades?",
  "How has The Simpsons addressed and incorporated shifts in media consumption, such as streaming services, into its distribution and content strategies?"
]

Report excerpt:

**Report: The Evolution of _The Simpsons_ (1989–Present)**
**1. Cultural Impact and Societal Relevance**
_The Simpsons_ debuted as a subversive critique of American middle‑class life, gaining notoriety for its bold satire in the 1990s... (truncated for brevity)

Complex director comparison sub‑queries:

["Who is the director of God's Gift To Women?", "Who is the director of Aldri annet enn bråk?", "What are the ages of the respective directors?", "Which director is older?"]

Retrieved director information:

The director of "God's Gift To Women" is Michael Curtiz, who was born on December 24, 1886 and passed away on April 11, 1962.
The director of "Aldri annet enn bråk" is Edith Carlmar, who was born on November 15, 1911 and passed away on May 17, 2003.
Therefore, Michael Curtiz is older than Edith Carlmar.