paperlib/AGENTS.md

AGENTS.md

Project overview

paperlib is a local-first paper library engine with a CLI.

It is designed to:

• import PDF papers into a structured local library
• convert PDFs into Markdown using external converters such as MinerU
• maintain stable per-paper metadata files and a searchable index database
• optionally generate AI-based structured summaries
• expose a clean CLI that is useful both for humans and for higher-level automation tools such as an arXiv daily digest workflow

paperlib is *not- primarily an AI app. AI summarization is an optional enrichment layer, not the core of the system.

The project should remain useful even when:

• no LLM API key is configured
• no summarization is enabled
• only import / convert / index / search features are used

---

Core design principles

1. Local-first

User data lives locally in the paper library directory.

The library must remain usable without a server, web app, or remote database.

Prefer plain files plus SQLite over opaque internal state.

2. CLI-first

The CLI is the primary interface.

All important workflows should be accessible from the CLI.

The Python API is useful, but secondary.

3. JSON files are the source of truth

Per-paper JSON files in the library are the durable source of truth.

Examples:

• meta.json
• summary.json

SQLite is an index/cache layer, not the canonical data store.

This means:

• the index must be rebuildable from files
• reindex should be able to repair the database from on-disk records
• code must not assume the database alone is authoritative

4. AI is optional enrichment

Importing, converting, indexing, listing, showing, and searching papers must work without AI.

AI summarization should be isolated behind a clean interface.

Do not make core workflows depend on an LLM provider.

5. Stable machine-readable interfaces

Important commands should support --json output so that other tools can consume them.

Examples:

• paperlib import ... --json
• paperlib summarize ... --json
• paperlib show ... --json
• paperlib export ... --format json

6. Small, explicit, inspectable components

Prefer simple and explicit logic over large hidden frameworks.

Keep components understandable:

• importer
• converter
• renderer
• summarizer
• search
• reindex
• doctor

Avoid unnecessary abstraction until there is a real need.

---

Non-goals

The following are currently out of scope unless explicitly planned later:

• mandatory daemon architecture
• web UI
• multi-user remote service
• cloud-first design
• vector database as a required dependency
• opaque agent framework controlling the core library
• “fully autonomous research assistant” behavior

---

Library data layout

The paper library on disk should be human-browsable.

A typical layout looks like:

library_root/
  config/
    config.toml
    vocab.yaml
    prompts/
      summarize_paper.md

  inbox/
  papers/
    arxiv/
      2026/
        2604.12345/
          meta.json
          source.pdf
          paper.md
          summary.json
          summary.md
          ref.bib
          assets/
          logs/
            mineru.log
    local/
      sha256-.../
        meta.json
        source.pdf
        paper.md
        summary.json
        summary.md

  db/
    paperlib.sqlite3

  cache/

Conventions:

• meta.json contains stable metadata and processing status
• summary.json contains structured AI-generated enrichment
• summary.md is rendered from summary.json
• paper.md is generated from the PDF by an external converter such as MinerU
• the database is rebuildable from the files above

---

Data model boundaries

meta.json

meta.json should contain deterministic or near-deterministic information, mostly from:

• import process
• file system state
• external paper metadata sources

Typical fields include:

• paper_id
• source_type
• source_id
• title
• authors
• published_date
• updated_date
• categories
• pdf_path
• paper_md_path
• summary_json_path
• summary_md_path
• imported_at
• conversion_status
• summary_status

Avoid putting speculative AI content into meta.json.

summary.json

summary.json is optional enrichment and may be regenerated.

It should contain structured fields such as:

• one-sentence summary
• problem statement
• method overview
• main results
• claimed contributions
• assumptions
• limitations
• problem tags
• technique tags
• entities
• relevance-to-user fields
• recommended sections

summary.json must include a schema version.

SQLite

SQLite stores searchable/indexed state and job-independent status.

It should help with:

• listing papers
• filtering and search
• path lookup
• tag lookup
• status overview

But it should never be treated as the only durable source of paper metadata.

---

CLI philosophy

The CLI should be easy for humans and predictable for scripts.

Important CLI expectations

• human-readable by default
• machine-readable with --json
• clear exit codes
• no hidden background magic
• no required daemon
• stable command names
• idempotent operations when possible

Expected command families

Core commands include:

• init
• import
• import-dir
• watch
• convert
• reindex
• doctor
• status
• list
• show
• search
• open
• print-path
• summarize
• render-summary
• export

When implementing commands, preserve a clear separation between:

• mutation commands
• read/query commands

---

Architecture guidelines

The codebase should be organized around a few clear layers.

1. Core domain logic

Pure Python logic for:

• identifying papers
• computing paths
• importing PDFs
• updating metadata
• converting PDFs to Markdown
• rendering summaries
• rebuilding the index

This layer should be testable without the CLI.

2. CLI layer

Thin wrappers around the core domain logic.

The CLI should:

• parse arguments
• call core functions
• format output
• handle exit codes

The CLI should not contain deep business logic.

3. Optional integrations

External systems should live in integration modules, for example:

• MinerU wrapper
• filesystem watch integration
• ripgrep integration
• LLM provider integration

Keep these adapters isolated.

4. Optional AI layer

The AI summarization layer should be behind a stable abstraction.

For example:

• load prompt template
• load paper markdown
• load optional profile / vocabulary
• call provider
• validate structured output
• write summary.json
• render summary.md

Avoid leaking provider-specific behavior into unrelated modules.

---

AI collaboration guidelines

When using AI to help develop this project, the AI should follow these rules.

1. Respect the project boundaries

Do not redesign paperlib into:

• a web app
• a required daemon
• a monolithic agent system
• a chat-first interface

Unless explicitly asked, keep the project aligned with:

• local-first
• CLI-first
• JSON/SQLite-based architecture
• AI-optional enrichment

2. Prefer incremental changes

Make small, reviewable changes.

When implementing a feature:

• first clarify which module owns it
• avoid broad refactors unless necessary
• preserve existing CLI semantics unless intentionally changing them

3. Keep file formats stable

Changes to meta.json or summary.json are important.

If changing schemas:

• update the schema version
• update documentation
• consider migration or backward compatibility
• do not silently break existing libraries

4. Avoid hidden coupling

Do not make unrelated modules depend on each other unnecessarily.

For example:

• search should not depend on LLM code
• import should not require summarization
• reindex should not assume a specific converter
• render-summary should not require calling AI again

5. Prefer explicit data flow

When adding features, keep data flow obvious.

For example:

• import creates or updates metadata
• convert creates paper.md
• summarize creates summary.json
• render-summary creates summary.md
• reindex rebuilds SQLite from files

6. Do not invent capabilities

If a feature is not implemented yet, do not pretend it exists.

Examples:

• do not write code that assumes a daemon exists
• do not assume remote sync exists
• do not assume vector search exists
• do not assume arXiv-specific logic belongs in the core library

7. Prefer durable outputs over polished prose

When designing AI summarization outputs, favor:

• structured JSON
• stable field names
• grep-friendly rendered Markdown
• concise, reusable information

over:

• highly polished review prose
• flashy but unstable output formats

---

Coding guidelines

General style

• Prefer straightforward Python.
• Use type hints.
• Keep functions small and focused.
• Add docstrings to public functions and classes.
• Avoid overengineering.
• Prefer composition over deep inheritance.

Error handling

• Fail clearly.
• Provide helpful error messages.
• Distinguish user-facing CLI errors from internal exceptions.
• Avoid silently swallowing errors.

Logging

• Use structured and informative logging where useful.
• Avoid noisy logs in normal CLI output.
• Keep machine-readable command output clean when --json is used.

File operations

• Be careful with moves, copies, and overwrites.
• Prefer atomic writes for JSON files when possible.
• Never corrupt existing metadata due to partial writes.

Idempotence

Where possible, commands should behave safely when run multiple times.

Examples:

• re-importing the same file should detect duplicates
• render-summary should be repeatable
• reindex should be safe to rerun

Testing

Add tests for:

• path layout logic
• metadata read/write behavior
• duplicate detection
• reindex behavior
• summary rendering
• search behavior
• CLI output contracts for core commands

Prefer unit tests for core logic and targeted integration tests for CLI behavior.

---

Search design guidelines

Search should support at least two useful modes:

1. Field-aware structured search

Examples:

• tags
• authors
• categories
• titles
• summary fields

2. Full-text-friendly search

Support grep-like workflows and integration with tools such as ripgrep.

Do not require semantic/vector search as a baseline feature.

If semantic search is ever added later, it should be optional and must not displace simple grep/database search.

---

Summarization design guidelines

Summarization should produce reusable structured outputs.

Summarization goals

A summary should be useful for:

• later human review
• grep-style reverse lookup
• building daily/weekly reports
• indexing by problem/method/result
• personal research triage

Summarization output

Prefer generating:

• summary.json as the canonical structured output
• summary.md rendered from JSON

Do not make free-form Markdown the only output.

Prompting guidelines

Prompts should instruct the model to:

• extract factual information
• avoid unsupported claims
• use concise and stable language
• prefer controlled vocabulary when available
• return structured JSON only
• use null or empty lists for unclear fields rather than hallucinating

Provider abstraction

The summarizer should not be tightly coupled to a single LLM provider.

Use a provider abstraction so the project can support:

• OpenAI-compatible APIs
• local models later if desired
• different prompt templates and vocabularies

---

What belongs in paperlib vs higher-level tools

paperlib is the base library engine.

It should own:

• PDF import
• local storage layout
• conversion to Markdown
• metadata files
• summary files
• index maintenance
• CLI access to those capabilities

It should not own high-level discovery workflows such as:

• arXiv daily fetching
• personalized new-paper ranking
• daily digest generation
• automated paper downloading from external feeds

Those belong in higher-level tools that consume paperlib.

---

Expected development workflow

When implementing a new feature, the preferred order is:

1. identify the owning module
2. define or update the data contract
3. implement the core logic
4. add tests
5. expose it through the CLI if appropriate
6. update docs and examples

If a change affects on-disk formats or CLI behavior, document it clearly.

---

Decision heuristics

When uncertain, prefer the option that is:

• more local-first
• more inspectable
• easier to test
• easier to recover from
• less coupled to AI
• more stable for scripts
• less magical

Examples:

• prefer JSON + Markdown over opaque internal blobs
• prefer explicit CLI commands over hidden automation
• prefer rebuildable indexes over fragile single-source databases
• prefer optional AI enrichment over mandatory AI workflows

---

Documentation expectations

Important features should be documented in:

• README.md for user-facing overview
• docs/cli.md for command behavior
• docs/storage-layout.md for on-disk structure
• docs/summary-schema.md for summary.json
• docs/integration-guide.md for higher-level tool integration

Keep docs aligned with actual behavior.

---

If you are an AI agent contributing code

Before making a change, ask:

1. Does this belong in paperlib, or in a higher-level workflow project?
2. Does this preserve local-first and CLI-first design?
3. Does this make AI optional, not mandatory?
4. Does this keep JSON files as the durable source of truth?
5. Does this keep the system understandable to a developer reading the code later?

If the answer to any of these is no, reconsider the approach.