How to Build a Document Research Database on Your Computer

The researcher's file problem

Over months or years, researchers accumulate documents. PDFs from journals, government reports, industry white papers, conference proceedings, internal analyses. They end up in nested folders with names like "Papers - March" and "Read Later."

The documents exist. The knowledge is there. But finding it again is the problem.

You remember reading something about supply chain resilience. Was it the McKinsey paper or the World Economic Forum brief? You open folders, scan filenames, skim first pages. Fifteen minutes later you either found it or gave up.

Your operating system can find a file named "supply-chain-report.pdf" but cannot tell you which of your 200 documents discusses supply chain resilience in the context of semiconductor manufacturing.

What you need is a structured, searchable index of what every document contains. Building that manually would take weeks. Building it with docrew takes an afternoon.

What the database looks like

The end product is a structured reference file containing the following for each source:

• Title: The actual document title, not the filename.
• Author(s): Who wrote or published it.
• Date: Publication or release date.
• Source/Publisher: The journal, organization, or institution.
• Document type: Academic paper, industry report, government publication, news article.
• Topic tags: Three to five keywords capturing the main subjects.
• Summary: A two-to-three-sentence description of what the document covers.
• Key findings: The main conclusions, arguments, or data points.
• Relevant quotes: Direct passages useful for reference.
• File path: Where the original lives on your machine.

This turns a folder of PDFs into something you can search, filter, sort, and cross-reference.

Step one: organize your source material

Consolidate your documents into your docrew workspace. You do not need a perfect folder structure -- the agent works recursively. But grouping by broad topic area helps:

• research/economics/
• research/technology/
• research/policy/
• research/industry-reports/

Supported formats include PDF, DOCX, and XLSX, all read natively by built-in parsers. For this walkthrough, assume roughly 200 documents across several subfolders.

Step two: initial indexing and metadata extraction

Open a new conversation in docrew and start the indexing task:

"Scan all documents in the research folder and its subfolders. For each document, extract the title, authors, publication date, source or publisher, and document type. Write the results to a CSV file called research-index.csv with one row per document. Include the file path as the first column."

The agent uses subagent delegation to process multiple files in parallel, keeping total time manageable -- typically 15 to 30 minutes for 200 documents, depending on length and complexity.

When the agent finishes, open the CSV and scan for accuracy. Titles should match, dates should be reasonable, author names should be correct. Fix any errors -- this index becomes the foundation for everything else.

Step three: generate summaries and key findings

With the basic index in place, go deeper:

"For each document in the research folder, write a two-to-three-sentence summary of the main topic and argument, and list the three to five most important findings. Add these as new columns to research-index.csv."

This pass requires the agent to read each document thoroughly -- not just scanning page one for metadata, but understanding the document well enough to summarize it.

For a 20-page industry report, the summary might read: "Analyzes semiconductor supply chain vulnerabilities following the 2024 chip shortage. Argues that geographic concentration in East Asia creates systemic risk. Recommends diversification of fabrication capacity."

Key findings might include: "Global semiconductor fabrication is 73% concentrated in Taiwan and South Korea. Lead time for new facilities is 3-5 years. Economic impact of a six-month supply disruption estimated at 490 billion dollars globally."

When you need to recall what a specific document said, read the summary instead of reopening a 50-page PDF.

Step four: add topic tags and cross-references

Give the database its search layer:

"Review the summaries and key findings for all documents in research-index.csv. Assign three to five topic tags from a consistent vocabulary: supply-chain, semiconductors, manufacturing, trade-policy, climate, energy, healthcare, finance, regulation, technology, AI, labor-markets, infrastructure, geopolitics. Add the tags as a new column. Then identify clusters of documents that address the same topic and note cross-references where findings support or contradict each other."

Using a consistent vocabulary makes filtering effective. You can search for "supply-chain" and get a focused set, rather than guessing between "logistics," "procurement," or "sourcing."

The cross-referencing step is where the database becomes more than the sum of its parts. The agent identifies connections you might not have noticed -- a 2024 academic paper about trade policy impacts on semiconductor availability relates directly to a 2025 industry report about manufacturing diversification.

Step five: extract relevant quotes

For research tasks requiring direct citations, extract quotes proactively:

"Go through each document and extract two to four notable direct quotes -- data points, strong arguments, or well-stated conclusions. Include the page number or section. Write to research-quotes.csv with columns for filename, quote text, page or section, and topic tag."

When writing a section about manufacturing costs, search the quotes file for "manufacturing" and find directly quotable passages, already extracted and attributed.

Keeping everything local

The entire database lives on your computer. Source documents, index CSV, quotes database -- all local files. Nothing is uploaded to a cloud service.

Confidentiality. Research collections often include pre-publication drafts and documents shared under NDA. A local database keeps them on your machine.

Permanence. Cloud services change terms, raise prices, or shut down. Your local database is a set of files that exist as long as you keep them.

Speed. Searching a local CSV is instantaneous. No network round-trip, no API call.

Portability. Copy your research folder to a backup drive or a new computer. The database is self-contained.

Using the database for research tasks

Literature review. Filter the index by relevant topic tags. Read summaries to identify which sources matter. Open only those documents.

Finding supporting evidence. Search the key findings column for relevant terms. The quotes database gives you ready-to-cite passages.

Identifying gaps. Sort by topic tag and date. All sources about a topic from 2024 or earlier? That gap tells you where to look for newer material.

Briefing others. Filter by topic, copy relevant summaries, and send them. A structured overview in minutes rather than days.

Trend analysis. Ask the agent: "Based on key findings across all documents tagged 'AI,' what are the three most frequently cited trends?" The agent reads extracted findings and identifies patterns across your entire collection.

Maintaining the database over time

When you add new documents, open a docrew conversation: "There are new files in the research folder not in research-index.csv. Identify them, extract the same metadata and summaries, and append them to the CSV."

The agent compares folder contents against the existing index, processes new files, and appends results. For five or ten new documents, this takes a couple of minutes.

Periodically, ask the agent to update cross-references and topic tags to incorporate new material. If your research focus shifts, re-tag older documents with an updated vocabulary or add new metadata fields.

Scaling up: beyond 200 documents

For larger collections, maintain one index per topic area or project rather than a single file. Process in batches of 50 to keep sessions focused.

The compound value of structured knowledge

The database is more valuable than any individual document in it. A single PDF contains one perspective. A structured database across 200 documents reveals patterns, contradictions, gaps, and connections.

The time investment is front-loaded. Building the initial database takes an afternoon. After that, each new document takes a minute to add. The return compounds as your collection grows and cross-references become richer.

docrew handles the mechanical work -- reading, extracting, summarizing, tagging, connecting. You handle the intellectual work -- deciding what matters and what the patterns mean.