Open your Zotero and take a look.
1,000 papers. All the PDFs are local, metadata is tidy, tags are in place. You’ve spent years building this library.
Now you need to write a review on large language models for code generation. You open Google Scholar and start searching.
Why not search in Zotero? Because you can’t remember what you’ve already downloaded. Zotero is a reference manager—it manages titles, authors, and years, not the content of papers. You can’t ask it “which papers discuss the effect of few-shot prompting on code generation?” It can only filter by metadata fields.
This is the problem every researcher with a large literature library runs into: you have a lot of material, but it might as well not exist.
Four Walls in the Traditional Workflow
Research work breaks down into roughly four stages, each with its own friction:
Literature search: You can’t remember what you’ve already downloaded, so you search Google Scholar and Semantic Scholar again. After finding a paper, you realize you downloaded it three months ago.
Reading and filtering: Even when you find a relevant paper, you don’t know which chapter covers the specific point you need. You open the PDF and ctrl+F, or scan from the beginning.
Note-taking: Notes are scattered across Notion, Zotero annotations, and various PDF markups—no unified view.
Writing the review: When making an argument, you know which papers support it, but you have to manually hunt down the original text to confirm the quote, page number, and logic.
Stack these four walls together, and literature review writing becomes painfully slow—not because you’re not working hard enough, but because of structural limitations in the tools themselves.
A Different Approach: Let AI Search Directly in Your Literature
The core idea behind Linkly AI is: instead of pre-filtering literature and handing it to AI, let AI browse your file cabinet freely, like a researcher would.
It builds a structured “card” (Outline Index) for each paper, letting AI progressively access your local literature library using three tools:
- search — search the literature, return a list of relevant files with abstracts
- outline — view the chapter structure of a paper, quickly locate the relevant section
- read — read the original text of a specific section
Through the MCP protocol, AI tools like Claude and Cursor can call these three tools directly. You just ask the question—the AI decides what to search for, what to read, and how much to read.
For the design principles behind this approach, see Outlines Index: A Progressive Disclosure Approach.
Setting Up Your Literature Library: 5 Minutes
Step 1: Export PDFs to a unified folder
In Zotero, you can bulk-export attachments or locate the storage directory directly (Zotero stores PDFs in ~/Zotero/storage/ by default, one subfolder per paper). We recommend organizing all PDFs into a single directory, such as ~/research-papers/.
Step 2: Build an index in Linkly AI
Download Linkly AI and install it. Open Linkly AI, select your literature folder, and start building the index. The following happen automatically in the background:
- BM25 full-text index (1–3 minutes; keyword search is available immediately after)
- Outline index (extracts chapter structure for each PDF)
- Vector semantic index (enables cross-language semantic search once complete)
Step 3: Configure MCP in your AI assistant
Follow the MCP configuration documentation—takes about 5 minutes. Once done, your AI assistant can access your local literature library.
Verify it’s working: Ask your AI assistant directly: “Search my computer for knowledge about the Transformer attention mechanism.” If it returns a list of relevant papers from your library, you’re good to go.
A Real Conversation Example
Here’s what the experience looks like once it’s configured.
You: Are there any papers in this library that discuss the limitations of Transformers in few-shot learning settings?
AI assistant (begins working):
First, searching for relevant papers:
search("transformer few-shot learning limitations")
Found 8 results:
#1 Few-Shot Learning with Transformers: A Survey
doc_id: 2847 | type: pdf | words: 18,400 | lines: 412
has_outline: yes | relevance: 0.94
snippet: "While transformers have demonstrated strong few-shot capabilities,
they face significant challenges including..."
#2 Meta-Learning for Low-Resource NLP Tasks
doc_id: 1923 | type: pdf | words: 9,200 | lines: 218
has_outline: yes | relevance: 0.87
snippet: "Transformer-based models require substantial prompt engineering
to achieve competitive few-shot performance..."
#3 Scaling Laws and Few-Shot Performance in Large Language Models
doc_id: 3341 | type: pdf | words: 22,100 | lines: 489
has_outline: yes | relevance: 0.81
snippet: "Despite impressive few-shot results, we identify three systematic
limitations in transformer architectures..."
... (5 more results with lower relevance)The top three are most relevant. Let me check their chapter structures:
outline(2847, 1923, 3341)
[doc_id: 2847] Few-Shot Learning with Transformers: A Survey
[2] Background [L18-67, 50 lines]
[2.1] Transformer Architecture Overview [L18-34]
[2.2] Few-Shot Learning Paradigms [L35-67]
[3] Capabilities and Benchmarks [L68-198, 131 lines]
[4] Limitations and Challenges [L199-312, 114 lines]
[4.1] Data Efficiency Constraints [L199-231]
[4.2] Distribution Shift Sensitivity [L232-271]
[4.3] Prompt Sensitivity [L272-312]
[5] Future Directions [L313-390, 78 lines]
[doc_id: 1923] Meta-Learning for Low-Resource NLP Tasks
...
[3] Transformer-Specific Challenges [L89-156, 68 lines]
[3.1] Attention Mechanism Limitations [L89-118]
[3.2] Token Representation in Low-Resource Settings [L119-156]
...
[doc_id: 3341] Scaling Laws and Few-Shot Performance...
...
[4] Systematic Limitations [L201-334, 134 lines]
[4.1] Memorization vs. Generalization [L201-245]
[4.2] Context Length Constraints [L246-289]
[4.3] Task Format Sensitivity [L290-334]
...Chapter 4 of the first paper and Chapter 4 of the third paper directly address limitations. Let me read those sections precisely:
read(2847, offset=199, limit=114)
read(3341, offset=201, limit=134)AI assistant (delivers the answer):
Based on three papers in your literature library, here are the main limitations of Transformers in few-shot learning settings:
1. Prompt Sensitivity
[Few-Shot Learning with Transformers, §4.3] finds that using different formats of few-shot examples for the same task produces accuracy differences of 15–30%. This means few-shot performance is highly dependent on prompt design and is difficult to reproduce reliably.
2. Data Efficiency Constraints
[Ibid., §4.1] notes that although Transformers can learn from a small number of examples, their data efficiency falls well short of methods specifically designed for few-shot learning (such as MAML), particularly in extreme low-resource scenarios (1–5 examples).
3. Distribution Shift Sensitivity
[Scaling Laws, §4.1] experiments show that Transformers in few-shot settings tend to rely more on memorized patterns from training data than on genuine generalization ability, leading to notable performance degradation on out-of-distribution test sets.
Would you like me to further compare the experimental setups across these three papers for these issues?
Throughout this process, the AI assistant autonomously made three tool calls: search → outline (three papers in parallel) → read (two target chapters). You asked one question; it returned an answer with precise citations.
Advanced Usage
Once you’re comfortable with the basics, you can do more complex things.
Cross-paper comparison
What do these 5 papers on RLHF have in common and where do they disagree
about its limitations? Cite specific sections from each.The AI assistant will call outline on each paper to locate the relevant sections on limitations, then read each one and synthesize a comparison.
Methodology extraction
Find all papers in my library that use contrastive learning,
and list their experimental setups including encoder architecture,
dataset, and batch size.The AI assistant will search first, then call outline on matching papers to find the experimental sections, then read them to extract the specific parameters.
Literature review drafting
Based on papers in my library about the evolution of attention mechanisms,
write an 800-word draft literature review organized chronologically,
with citations for each claim.This is the most valuable use case. The AI assistant will actively search, read, and synthesize, giving you a draft with real citations—not hallucinated references based on general knowledge.
An Important Note
The core value of this workflow is: the AI assistant reads literature that actually exists in your library, and cites original text from your local files. It won’t fabricate citations because it has real text to read from.
This is fundamentally different from just asking an AI assistant to write a literature review for you—the latter often produces hallucinated references because there’s no original text to verify against.
That said, the AI assistant’s summaries and syntheses can still have errors. Before finalizing any citation, it’s always worth returning to the original text to verify key claims.
Beyond Academic Research
One last thing worth mentioning: this workflow isn’t just for academic literature.
It applies to any situation where you need to manage a large corpus of reference material:
- Lawyers: case libraries, legal texts, contract templates
- Journalists: interview transcripts, industry reports, background material
- Consultants: industry research reports, client history, methodology documents
- Product managers: competitive analysis, user research reports, internal documentation
At its core, this workflow solves the same problem in every domain: you have a large body of existing material, but every time you need something, you’re starting from scratch.
Get started: Quickstart — from installation to your first search, about 10 minutes.
Made by the Linkly AI team.