AI mode documentation (#2620)

docs/source/command-line-options.rst

@@ -161,6 +161,14 @@ ndjson (one report per username). E.g. ``--json ndjson``
 ``-M``, ``--md`` - Generate a Markdown report (general report on all
 usernames). See :ref:`markdown-report` below.
 
+``--ai`` - Run an AI-powered analysis of the search results using an
+OpenAI-compatible chat completion API. The internal Markdown report is
+sent to the model, which returns a short investigation summary that is
+streamed to the terminal. See :ref:`ai-analysis` below.
+
+``--ai-model`` - Model name to use with ``--ai``. Defaults to
+``openai_model`` from settings (``gpt-4o`` out of the box).
+
 ``-fo``, ``--folderoutput`` - Results will be saved to this folder,
 ``results`` by default. Will be created if doesn’t exist.
 
@@ -242,3 +250,51 @@ The Markdown format is optimized for LLM context windows. You can feed the repor
 
 The structured Markdown with per-site sections makes it easy for AI tools to extract relationships, cross-reference identities, and identify patterns across accounts.
 
+For a built-in alternative that calls the model for you and prints the
+summary directly, see :ref:`ai-analysis` below.
+
+.. _ai-analysis:
+
+AI analysis (built-in)
+----------------------
+
+The ``--ai`` flag turns the search results into a short investigation
+summary by sending the internal Markdown report to an OpenAI-compatible
+chat completion API and streaming the model's reply to the terminal.
+
+.. code-block:: console
+
+   export OPENAI_API_KEY=sk-...
+   maigret username --ai
+
+   # use a smaller / cheaper model
+   maigret username --ai --ai-model gpt-4o-mini
+
+While ``--ai`` is active, per-site progress lines and the short text
+report at the end are suppressed so the streamed summary is the main
+output. The Markdown report itself is built in memory and is **not**
+written to disk by ``--ai`` alone — combine with ``--md`` if you also
+want the file on disk.
+
+The summary follows a fixed format with sections for the most likely
+real name, location, occupation, interests, languages, main website,
+username variants, number of platforms, active years, a confidence
+rating, and a short list of follow-up leads. The model is instructed
+to rely only on what is supported by the report and to avoid mixing
+clearly unrelated profiles into the main identity.
+
+**Configuration.** The API key is resolved from
+``settings.openai_api_key`` first, then from the ``OPENAI_API_KEY``
+environment variable. The endpoint defaults to
+``https://api.openai.com/v1`` and can be redirected to any
+OpenAI-compatible service (Azure OpenAI, OpenRouter, a local server,
+…) by setting ``openai_api_base_url`` in ``settings.json``. See
+:ref:`settings` for the full list of options.
+
+.. note::
+
+   ``--ai`` makes a network request to the configured chat completion
+   endpoint and sends the full Markdown report (which contains the
+   gathered profile data). Use it only with providers and accounts
+   you trust with that data.
+

docs/source/features.rst

@@ -147,6 +147,33 @@ Also, there is a short text report in the CLI output after the end of a searchin
 .. warning::
    XMind 8 mindmaps are incompatible with XMind 2022!
 
+AI analysis
+-----------
+
+Maigret can produce a short, human-readable investigation summary on top
+of the raw search results using the ``--ai`` flag. It builds the
+internal Markdown report, sends it to an OpenAI-compatible chat
+completion endpoint, and streams the model's reply directly to the
+terminal.
+
+.. code-block:: console
+
+   export OPENAI_API_KEY=sk-...
+   maigret username --ai
+
+The summary uses a fixed format with the most likely real name,
+location, occupation, interests, languages, main website, username
+variants, number of platforms, active years, a confidence rating, and a
+short list of follow-up leads. While ``--ai`` is active, per-site
+progress and the short text report are suppressed so the streamed
+summary is the main output.
+
+The endpoint, model, and API key are configured via ``settings.json``
+(``openai_api_key``, ``openai_model``, ``openai_api_base_url``) or the
+``OPENAI_API_KEY`` environment variable. Any OpenAI-compatible API can
+be used (Azure OpenAI, OpenRouter, a local server, …). See
+:ref:`ai-analysis` and :ref:`settings` for details.
+
 Tags
 ----
 

docs/source/settings.rst

@@ -101,3 +101,51 @@ This is recommended for **Docker containers**, **CI pipelines**, and **air-gappe
      - URL of the metadata file (for custom mirrors)
 
 **Using a custom database** with ``--db`` always skips auto-update — you are explicitly choosing your data source.
+
+.. _ai-analysis-settings:
+
+AI analysis
+-----------
+
+The ``--ai`` flag (see :ref:`ai-analysis`) talks to an OpenAI-compatible
+chat completion API. Three settings control how that request is made:
+
+.. list-table::
+   :header-rows: 1
+   :widths: 35 25 40
+
+   * - Setting
+     - Default
+     - Description
+   * - ``openai_api_key``
+     - ``""`` (empty)
+     - API key. If empty, Maigret falls back to the ``OPENAI_API_KEY``
+       environment variable.
+   * - ``openai_model``
+     - ``gpt-4o``
+     - Default model name. Overridable per-run with ``--ai-model``.
+   * - ``openai_api_base_url``
+     - ``https://api.openai.com/v1``
+     - Base URL of the chat completion API. Point this at any
+       OpenAI-compatible service (Azure OpenAI, OpenRouter, a local
+       server, …) to use it instead of OpenAI directly.
+
+Example ``~/.maigret/settings.json`` snippet using a non-OpenAI
+endpoint:
+
+.. code-block:: json
+
+   {
+       "openai_api_key": "sk-...",
+       "openai_model": "gpt-4o-mini",
+       "openai_api_base_url": "https://openrouter.ai/api/v1"
+   }
+
+The key resolution order is ``settings.openai_api_key`` → ``OPENAI_API_KEY``
+environment variable; the first non-empty value wins.
+
+.. note::
+
+   ``--ai`` sends the full internal Markdown report (which contains the
+   gathered profile data) to the configured endpoint. Only use providers
+   and accounts you trust with that data.