Question 1

How is this different from a search wrapper around an LLM?

Accepted Answer

The program records real tool calls, sources, evidence, and visible context instead of trusting the model narrative. State, checkpoints, and citation checks are separate from chat history so runs can be inspected and resumed.

Question 2

What is the difference between checkpoints and long-term memory?

Accepted Answer

They are different stores. SQLite checkpoints save one task so it can resume by thread_id after interruption, while memory saves reusable facts for later tasks and only enters the request as memory context.

Question 3

Can I try it before buying model access?

Accepted Answer

Start with the fully offline demo, which needs no API key, then copy the env template and point ChatOpenAI-compatible settings at your model service. Real search and report runs need a key and network access.

Question 4

Does a finished report with citations mean the facts are right?

Accepted Answer

No. The project warns that having citations does not mean facts are correct and probability scores do not replace human verification. Budgets and forced收尾 can also finish a report from incomplete evidence.

Question 5

How do I run it, and what does it need?

Accepted Answer

Python 3.12 or newer, with uv recommended. Run uv sync --extra dev --extra web, then uv run deepresearch demo works fully offline with no API key, and the suite is 332 automated tests. For real runs, copy .env.example and point the ChatOpenAI-compatible settings at a provider; DeepSeek is the worked example, defaulting to DEEPRESEARCH_API_KEY, DEEPRESEARCH_MODEL and DEEPRESEARCH_BASE_URL.

Question 6

What is the Skills layer, and how is it kept from running away?

Accepted Answer

Tools answer what the agent can do; Skills answer how to complete a class of task, and they ship as SKILL.md files under a skills directory you point at with environment variables. Controlled evolution is the safety property: cross-task evaluation generates candidates, rules and a probabilistic review pass over them, and a human promotes a version. Older versions are retained for rollback, so the model never edits a live skill on its own.

Question 7

How does it stop a run that will not end?

Accepted Answer

With token governance, search and reading budgets, bounded reflection and a forced wrap-up stage. The README states the cost plainly: when a budget boundary is reached, the report may be generated from incomplete evidence. Run a real task with --stream --show-trace and --output reports/name.md, and note the program will not overwrite a file that already exists.

Deepreseach Learing

Highlights

Questions about this entry

More Jev resources like this

From the Jev community