Agentic Chat

Agentic Chat is Rankion's natural-language interface to the entire platform. Instead of switching between 40+ tools, you write a question. The master agent asks for clarification when ambiguous, builds a structured plan, delegates steps to specialized subagents, calls tools, observes results, and returns a summarized answer. All plans are persisted and queryable, and all tool calls are billed in credits the same way they would be otherwise. For power users who'd rather reach platform data by question than by click.

What it can do

• Brainstorm → Plan → Execute pipeline — every request runs through the same pipeline. Trivial = 1-step plan, complex = multi-step with subagents. Zero approval gates.
• Clarification questions — when ambiguous, the agent triggers clarify_intent with 1–3 questions instead of guessing.
• Subagent delegation — steps are dispatched to Analyze, Research, or Action subagents that run their own tool calls and report back via report_to_master.
• Tool calling — access to all modules: score AI visibility, research keywords, generate articles, pull backlinks, check SERP positions, create action items, send emails, analyze images, query your own data. Data queries only read your own team's data and never change anything.
• Live UI with SSE — sticky live pill while the plan runs, permanent card in the stream after it finishes, opt-in sidebar for drill-down on every step including output JSON and tool calls.
• Persistent plans — dedicated tables agentic_plans (1 row/plan) and agentic_plan_steps (N rows/plan), plan history per session, smart retry with invalidation of dependent steps.
• Wiki integration — the query_wiki tool searches this wiki and cites sources.
• Suggestions that fit your data — the tiles above the input field are based on what your team has already created. A suggestion only appears when the data for it exists: “5 content ideas for your most important topic area”, for example, only once you have keywords or articles — otherwise you will see a starter suggestion like “Create first project”. That way, no suggestion ever costs you credits for an answer that, for lack of data, cannot be one.

Example prompts

• "Which of my keywords don't rank in the top 10?"
• "Write an article about replacing shock absorbers and publish it to blog autohaus-mueller."
• "Who is cited in ChatGPT for vegan sneakers that isn't us?"
• "Which backlinks did we lose in the last 30 days and are toxic?"
• "Create a Reality Check PDF from the latest tracking run and email it to info@firma.de."
• "Build me a cluster of 50 long-tails for crm software with KD<35."

When to use

• You don't know the click path in the UI or want to skip it.
• You need an answer that combines data from 3+ modules (e.g. backlinks × Rank Tracker × AVI).
• You want a repeatable mini-pipeline without code (Generate → Score → Publish).
• You want plan telemetry (which step, which tools, which credits) as an audit trail.

Workflow

1. Open a session — UI at /chat or via API. Optional gallery_image_id for visual queries.
2. Ask — POST /agentic/chat with {message, session_id?}.
3. Answer clarifications — when the agent asks back, just reply normally in the next message.
4. Watch the plan — the live pill shows steps + current status. Sidebar for drill-down.
5. Get the result — the response contains text, tool_calls[], iterations, credits_used, duration_ms, resumed.

For async pulls (e.g. DataForSEO), the endpoint waits server-side via wait_for_final=true (default) for up to ~480 s for the final resume turn — resumed: true means the text is the finished answer, not the wait notice.

API

Body of /agentic/chat:

{
  "message": "Which backlinks did I lose in 30 days?",
  "session_id": 1234
}

Credits & Limits

• Variable. Credits depend on the tools the agent calls (e.g. Keyword Research = 5, Generate = 5, Image Analyze = depends on mode).
• Iterations cap: 150 tool calls per turn (env-overridable).
• Wall-clock cap: 25 minutes per turn.
• No-progress detector: 8 — on stall, the agent switches to a tools-disabled summarize call instead of a hard error.
• PHP-FPM hard limit: 600 s synchronous response. Set the client timeout accordingly.
• Triple-scoped: team_id + user_id + session_id against IDOR.

Ask a follow-up about part of the answer

Chat answers are often long, and your follow-up question usually concerns just one sentence from it. Instead of retyping it, select it: as soon as you select text in a chat message, a small Ask AI button appears next to it.

Clicking it places the selected text as a quote before your draft in the input field — as a Markdown quote, i.e. every line preceded by >. The cursor jumps to the end, and the input field prompts you with "Your question about the selected text …" for the actual question. You type it after that and send.

The quote is the context. It's sent as part of your message, not as invisible extra information — so you see exactly what the AI sees, and in your message bubble the quote later stays visually set apart.

Three things about the button:

• It only appears in a chat message. If you select text in the header or in the empty state, nothing happens — that's intentional, otherwise it would be in the way everywhere.
• It disappears on its own when you click elsewhere, clear the selection, scroll, or press Escape.
• It's also reachable without a mouse. If you select with Shift and the arrow keys, the button only gets focus once you release the Shift key — as long as you hold it, you can keep extending the selection. On mobile, the button appears below the selection so your finger doesn't cover it.

Stop a running request

While "Rankion AI denkt nach…" is shown, a Stop button sits next to it. Clicking it ends the request: the AI stops working, and you can immediately ask something new — you don't have to wait for the old run to finish on its own.

Three things worth knowing:

• No further costs after stopping. What the AI had already done by then has happened, though, and stays billed — stopping doesn't refund credits already spent. The earlier you stop, the less is used.
• A response already in progress stays visible. If the AI had already written text, it stays in the history, marked as incomplete. So you can see what you paid for and build on it.
• A running plan is stopped along with it. If the AI had mapped out several steps for your question, those are cancelled too — nothing keeps running in the background.

Stopping is also reachable via the API: POST /v1/agentic/sessions/{session}/abort, at no credit cost.

File downloads (generate_download)

When you ask the agent for a file ("give me the plan as Markdown/Excel/CSV"), it uses the generate_download tool and returns a signed download URL (valid 24 h, stored under your account, auto-deleted after 30 days).

• The full file content goes into the required content field — for md/txt/json the complete text/document, for xlsx/csv an array of row objects. The description field is only a short note next to the link, never the content itself.
• Background: the agent temporarily confused content and description, putting the report text into description → the tool rejected with "content ist Pflicht" and the turn ran into the no-progress detector. Fixed by a sharpened tool description (tool: App\Agentic\Tools\GenerateDownloadTool). Required fields: format, filename, content.

Related modules

• AI Visibility Tracking — the agent can start tracking runs and interpret AVI values.
• Keyword Explorer & Research — research tools are directly callable.
• AI Content Editor — Generate, Score, Optimize, Publish fully orchestratable.
• Backlinks Module — backlink pulls and toxic detection by question.