🤖 AI

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

This paper proposes Direct Corpus Interaction (DCI), a retrieval paradigm where agents bypass fixed semantic interfaces to interact directly with raw corpora using general-purpose tools, demonstrating that this approach significantly outperforms traditional lexical and semantic retrievers on complex agentic search tasks by enabling flexible, multi-step evidence exploration without the need for offline indexing.

Original authors: Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, Yuyang Bai, Shangbin Feng, Hangxiao Zhu, Ming Zhong, Yuyu Zhang, Jianwen Xie, Yejin Choi, James Zou, Jiawei Han, Wenhu Chen, Jimmy Lin

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, Yuyang Bai, Shangbin Feng, Hangxiao Zhu, Ming Zhong, Yuyu Zhang, Jianwen Xie, Yejin Choi, James Zou, Jiawei Han, Wenhu Chen, Jimmy Lin, Dongfu Jiang, Yu Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a complex mystery. You have a massive library of books (the "corpus") containing all the clues you need.

The Old Way: The Librarian with a Filter
Traditionally, when a detective (an AI agent) needs a clue, they ask a librarian (a "retriever") for help. The detective says, "I'm looking for something about a stolen diamond," and the librarian uses a special scanner to find the top 5 books that sound like they might be about diamonds. The librarian hands over just those 5 pages.

The problem? If the clue is actually hidden in a sentence on page 400 of a book the librarian didn't pick, the detective never sees it. Also, if the detective needs to check if a specific word appears exactly as written, or combine two weak clues ("find the book with 'diamond' AND 'red' AND '2024'"), the librarian's pre-filtered list often fails. The detective is stuck with what the librarian gave them, even if the librarian missed the real answer.

The New Way: The Detective with a Flashlight
This paper proposes a new method called Direct Corpus Interaction (DCI). Instead of asking a librarian to filter the books, the detective is given a flashlight and a key, and they walk into the library themselves.

They don't use a fancy scanner. Instead, they use simple, powerful tools they already know:

grep: A tool to shout "Find every page that says 'diamond'!" and instantly get every single match.
find: A tool to locate exactly which shelf a book is on.
head/tail: Tools to peek at just the beginning or end of a page without reading the whole thing.

Why is this better?
The authors argue that modern AI agents are getting smart enough to be their own librarians. When the agent can walk into the library and use these simple tools, they can:

Be Exact: They can demand, "Show me only the pages where 'diamond' and 'red' appear in the same sentence." A standard librarian might miss this if the words are far apart.
Follow the Trail: If they find a clue, they can immediately check the surrounding text to see if it connects to another clue, without waiting for a new "search" request.
Avoid the "Top 5" Trap: They aren't limited to the top 5 results. If the answer is in the 500th book, they can find it.

The Results
The researchers tested this "flashlight" method against the "librarian" method on several difficult tasks:

Deep Research: On a test called BrowseComp-Plus, where agents had to solve complex riddles using a closed library of text, the flashlight method got 80% of the answers right, while the best librarian method only got 69%. It also cost less money to run.
Multi-Step Questions: On questions requiring multiple steps of logic (like "Who is the brother of the person who won the award in 2020?"), the flashlight method scored 83%, crushing the previous best method which scored only 52%.
Ranking: Even when the task was just to rank documents by relevance, the flashlight method was significantly more accurate.

The "Resolution" Concept
The paper introduces a cool concept called Resolution.

Low Resolution (The Librarian): You see the whole book as a blurry blob. You know it's about the topic, but you can't see the fine details.
High Resolution (The Flashlight): You can zoom in on a single word, a specific sentence, or a specific paragraph. You can verify the exact spelling and context.

The authors found that the flashlight method won not necessarily because it found more books, but because once it found a relevant book, it could zoom in and extract the exact, tiny piece of evidence needed to solve the puzzle. The librarian method often found the right book but couldn't "see" the specific clue inside it clearly enough.

The Catch
There is a trade-off. If the library is tiny, the librarian is fast. But if the library is huge and the detective has to search through millions of files, the flashlight method can get slow and expensive because the agent has to do more work to find the needle in the haystack. However, once the agent finds a promising file, the flashlight method is unbeatable at digging out the truth.

In Summary
This paper suggests that for smart AI agents, we should stop treating the "search" step as a black box that hands over a pre-filtered list. Instead, we should give the agent direct, high-resolution access to the raw data using simple command-line tools. This allows the agent to act more like a human researcher: hunting for clues, checking exact details, and piecing together evidence step-by-step, rather than just hoping the librarian picked the right pages.

Technical Summary: Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

Problem Statement

Modern retrieval systems, whether lexical (e.g., BM25) or semantic (e.g., dense embeddings), rely on a fixed similarity interface that compresses access to a corpus into a single top- $k$ retrieval step prior to reasoning. While efficient, the authors argue this abstraction creates a bottleneck for agentic search. Complex agentic tasks—such as those found in benchmarks like BrowseComp-Plus—require agents to orchestrate multi-step processes involving exact lexical constraints, sparse clue conjunctions, local context verification, and hypothesis refinement.

In standard retrieval-augmented pipelines, evidence filtered out early by a retriever cannot be recovered by downstream reasoning, even if the agent possesses strong reasoning capabilities. Furthermore, the fixed top- $k$ interface prevents agents from performing fine-grained operations like chaining regex matches, verifying exact spans, or iteratively refining search plans based on partial evidence. As agents become more capable, the limitation shifts from the model's reasoning ability to the resolution of the interface through which it interacts with the corpus.

Methodology: Direct Corpus Interaction (DCI)

The paper proposes Direct Corpus Interaction (DCI) as a new retrieval paradigm. Instead of querying a pre-built index or a semantic retriever API, the agent interacts directly with the raw corpus using general-purpose terminal tools (e.g., grep, find, read, bash, and lightweight scripts).

Core Mechanics

Interface: The agent bypasses offline indexing, vector databases, and embedding models. It treats the corpus as a file system accessible via a command-line interface (CLI).
Operations: The agent composes operations to:
- Explore: Navigate directory structures (find, ls).
- Match: Perform exact or regular-expression matching (grep, rg).
- Inspect: Read local context around matches (head, tail, sed, read).
- Compose: Chain commands to enforce complex constraints (e.g., grep 'foo' file | grep 'bar').
Implementation: The authors instantiate DCI using two agent scaffolds:
- DCI-Agent-Lite: A minimal harness based on Pi, restricted to raw terminal interaction (bash and read) with a lightweight context management layer (truncation, compaction, summarization) to handle long-horizon search.
- DCI-Agent-CC: A stronger instantiation using Claude Code as the CLI agent, providing robust tool orchestration and context handling.

Evaluation Metrics

Beyond standard accuracy, the authors introduce Retrieval Interface Resolution as a conceptual lens, measured by:

Coverage: Whether the trajectory surfaces relevant (gold) documents.
Localization: How efficiently the trajectory narrows to a small, usable evidence span within a surfaced document. This metric quantifies the ability to extract concentrated evidence rather than just retrieving a document.

Key Contributions

Formalization of DCI: The paper formalizes Direct Corpus Interaction as a retrieval paradigm, systematically evaluating it across agentic search, multi-hop QA, and IR ranking tasks without relying on offline indices.
Performance Superiority: DCI demonstrates competitive or superior performance to strong sparse, dense, and reranking baselines across diverse benchmarks, often without requiring external retrievers.
Resolution Lens: The authors introduce "retrieval interface resolution" to explain DCI's effectiveness. They argue that for capable agents, the interface's granularity (ability to operate on units smaller than entire documents) is as critical as the retriever's ranking quality.

Experimental Results

The authors evaluated DCI on three benchmark families: BrowseComp-Plus (agentic search), Multi-hop QA (NQ, TriviaQA, HotpotQA, etc.), and IR Ranking (BRIGHT, BEIR).

Agentic Search (BrowseComp-Plus): Replacing a Qwen3-Embedding-8B retriever with DCI (using the same Claude Sonnet 4.6 backbone) improved accuracy from 69.0% to 80.0% (+11.0 points) while reducing API costs by 29.4%. DCI-Agent-CC outperformed the strongest retrieval baseline (GPT-5 + Qwen3-Embedding-8B) by 8.3 points.
Multi-hop QA: DCI-Agent-CC achieved 83.0% average accuracy, surpassing the strongest retrieval-agent baseline (ASearcher-Local-14B) by 30.7 points. DCI-Agent-Lite also achieved 68.0%, significantly outperforming retrieval baselines.
IR Ranking: On BRIGHT and BEIR datasets, DCI-Agent-CC achieved an average NDCG@10 of 68.5, outperforming the best retrieval baseline (ReasonRank-32B) by 21.5 points.

Mechanism Analysis

Trajectory analysis reveals that DCI's gains do not primarily stem from recovering more gold documents (coverage is often lower than dense retrievers). Instead, the advantage comes from fine-grained discovery and verification:

DCI agents often succeed where retrieval agents fail because they can convert surfaced evidence into high-value local search steps (e.g., verifying exact string matches or chaining weak clues).
The "resolution" of the interface allows agents to pivot from broad retrieval to precise inspection once a relevant document is found.
Scaling: DCI scales well with search depth but incurs rapidly rising costs with search breadth (corpus size), suggesting an operating envelope where it excels in focused, deep exploration rather than broad, shallow scanning.

Significance and Claims

The paper claims that as language agents become stronger, retrieval quality depends not only on reasoning ability but on the resolution of the interface. The authors argue that retrieval for capable agents should be reframed as an interface-design problem rather than solely a retriever-design problem.

Paradigm Shift: DCI suggests that compressed similarity indexes can become bottlenecks when agents need to search like human researchers (formulating hypotheses, testing exact patterns, reading local context).
Practicality: DCI requires no offline embedding or indexing, adapts naturally to evolving local corpora, and allows agents to operate directly within their reasoning environment.
Design Space: Dense and sparse retrieval remain effective for large, static corpora, but DCI occupies a distinct point in the broader design space of corpus interfaces, offering high-resolution access for tasks requiring precise, multi-step evidence aggregation.

The authors conclude that DCI is not merely another approach but evidence that the future of agentic search lies in designing interfaces that align with the agent's reasoning capabilities, enabling direct, high-resolution interaction with raw data.