Agentic Paper Digest

When to use

• - Fetch a recent paper digest from arXiv and Hugging Face.
• Produce JSON output for downstream agents.
• Run a local API server when a polling workflow is needed.

Prereqs

• - Python 3 and network access.
• LLM access via OPENAI_API_KEY or an OpenAI-compatible provider via LITELLM_API_BASE + LITELLM_API_KEY.
• INLINECODE3 is optional for bootstrap; otherwise curl/wget (or Python) is used to download the repo.

Get the code and install

• - Preferred: run the bootstrap helper script. It uses git when available or falls back to a zip download.

CODEBLOCK0

• - Override the clone location by setting PROJECT_DIR.

CODEBLOCK1

Run (CLI preferred)

CODEBLOCK2

• - Pass through CLI flags as needed.

CODEBLOCK3

Run (API optional)

CODEBLOCK4

• - Trigger runs and read results.

CODEBLOCK5

• - Stop the API server if needed.

CODEBLOCK6

Outputs

• - CLI --json prints run_id, seen, kept, window_start, and window_end.
• Data store: data/papers.sqlite3 (under PROJECT_DIR).
• API: POST /api/run, GET /api/status, GET /api/papers, GET/POST /api/topics, GET/POST /api/settings.

Configuration

Config files live in PROJECT_DIR/config. Environment variables can be set in the shell or via a .env file. The wrappers here auto-load .env from PROJECT_DIR (override with ENV_FILE=/path/to/.env).

Environment (.env or exported vars)

• - OPENAI_API_KEY: required for OpenAI models (litellm reads this).
• INLINECODE26, LITELLM_API_KEY: use an OpenAI-compatible proxy/provider.
• INLINECODE28, LITELLM_MODEL_SUMMARY: models for relevance and summarization (summary defaults to relevance model if unset).
• INLINECODE30, LITELLM_TEMPERATURE_SUMMARY: lower for more deterministic output.
• INLINECODE32: retry count for LLM calls.
• INLINECODE33: drop unsupported params to avoid provider errors.
• INLINECODE34, APP_TZ: recency window and timezone.
• INLINECODE36: comma-separated categories (default includes cs.CL,cs.AI,cs.LG,stat.ML,cs.CR).
• INLINECODE38, HF_API_BASE: override source endpoints if needed.
• INLINECODE40, ARXIV_PAGE_SIZE: arXiv paging limits.
• INLINECODE42: cap candidates per source before LLM filtering.
• INLINECODE43, REQUEST_TIMEOUT_S: source fetch and per-request timeouts.
• INLINECODE45: include first-page PDF text in summaries; requires PyMuPDF (pip install pymupdf).
• INLINECODE48: location for papers.sqlite3.
• INLINECODE50: comma-separated origins allowed by the API server (UI use).
• Path overrides: TOPICS_PATH, SETTINGS_PATH, AFFILIATION_BOOSTS_PATH.

Config files

• - config/topics.json: list of topics with id, label, description, max_per_topic, and keywords. The relevance classifier must output topic IDs exactly as defined here. max_per_topic also caps results in GET /api/papers when apply_topic_caps=1.
• INLINECODE63: overrides fetch limits (arxiv_max_results, arxiv_page_size, fetch_timeout_s, max_candidates_per_source). Updated via POST /api/settings.
• INLINECODE69: list of {pattern, weight} boosts applied by substring match over affiliations. Weights add up and are capped at 1.0. Invalid JSON disables boosts, so keep the file strict JSON (no trailing commas).

Mandatory workflow (follow step-by-step)

1. 1. You first MUST open and read the configuration from the github repo: https://github.com/matanle51/agenticpaperdigest you downloaded:

- Load config/topics.json, config/settings.json, and config/affiliations.json (if present). - Note current topic IDs, caps, and fetch limits before asking the user to change them.

1. 2. ASK THE USER TO PROVIDE IT'S PREFERENCES ABOUT THE FOLLOWING (HELP THE USER):

- Topics of interest → update config/topics.json (topics[].id/label/description/keywords, max_per_topic). Show current defaults and ask whether to keep or change them. - Time window (hours) → set WINDOW_HOURS (or pass --window-hours to CLI) only if the user cares; otherwise keep default to 24h. - ASK THE USER TO FILL THE FOLLOWING PARAMETERS (explain the user why are their intent): ARXIV_CATEGORIES, ARXIV_MAX_RESULTS, ARXIV_PAGE_SIZE, MAX_CANDIDATES_PER_SOURCE. Ask whether to keep defaults and show the current values. - Model/provider → set OPENAI_API_KEY or LITELLM_API_KEY (+ LITELLM_API_BASE if proxy), and set LITELLM_MODEL_RELEVANCE/LITELLM_MODEL_SUMMARY. - Do NOT ask by default: timezone, quality vs cost, timeouts, PDF text, affiliation biasing, sources list. Use defaults unless the user requests changes.

1. 3. Confirm workspace path: Ask where to clone/run. Default to PROJECT_DIR="$HOME/agentic_paper_digest" if the user doesn’t care. Never hardcode /Users/... paths.
2. Bootstrap the repo: Run the bootstrap script (unless the repo already exists and the user says to skip).
3. Create or verify .env:

- If .env is missing, create it from .env.example (in the repo), then ask the user to fill keys and any requested preferences. - Ensure at least one of OPENAI_API_KEY or LITELLM_API_KEY is set before running.

1. 6. Apply config changes:

- Edit JSON files directly (or use POST /api/topics and POST /api/settings if running the API).

1. 7. Run the pipeline:

- Prefer scripts/run_cli.sh for one-off JSON output. - Use scripts/run_api.sh only if the user explicitly asks for UI/API access or polling.

1. 8. Report results:

- If results are sparse, suggest increasing WINDOW_HOURS, ARXIV_MAX_RESULTS, or broadening topics.

Getting good results

• - Help the user define and keep topics focused and mutually exclusive so the classifier can choose the right IDs.
• Use a stronger model for summaries than for relevance if quality matters.
• If using openAI's model, defualy to gpt-5-mini for good tradeoff.
• Increase WINDOW_HOURS or ARXIV_MAX_RESULTS when results are sparse, or lower them if results are too noisy.
• Tune ARXIV_CATEGORIES to your research domains.
• Enable PDF text (ENABLE_PDF_TEXT=1) when abstracts are too thin.
• Use modest affiliation weights to bias ranking without swamping relevance.
• BE PROACTIVE AND HELP THE USER TUNE THE SKILL FOR GOOD RESULTS!

Troubleshooting

• - Port 8000 busy: run bash "{baseDir}/scripts/stop_api.sh" or pass --port to the API command.
• Empty results: increase WINDOW_HOURS or verify the API key in .env.
• Missing API key errors: export OPENAI_API_KEY or LITELLM_API_KEY in the shell before running.