Research Pipeline Workflow

Research pipeline workflow

Orchestrates a deployed Mistral Workflow for multi-step research (literature review, competitive analysis, technical deep-dive). Queries hypotheses at each checkpoint, lets the user validate or amend them, and injects additional sources when needed — all via workflow_interact.

Profile note: workflow_execute, workflow_status, and workflow_interact are available in the core profile (default). No additional profile is required for workflow-only orchestration.

Temporal behavior: the workflow status stays RUNNING even when the pipeline is blocked waiting for hypothesis validation (a wait_for_input() pause in the workflow). Detect this state by querying a status handler — do not wait for a PAUSED status, which does not exist in this API.

Important: handler names are workflow-specific

Ask the user (or consult mistral://workflows) for:

• Query handler for progress: e.g. "get_progress"
• Query handler for hypotheses: e.g. "get_hypotheses"
• Signal handler for hypothesis decisions: e.g. "hypothesis_decision"

Steps

Step 1 — Define the research mission

Ask the user for:

1. workflowIdentifier — the deployed research workflow name or ID
2. topic — the research question (be specific)
3. depth — "shallow" (quick scan), "medium" (balanced), "deep" (comprehensive)
4. output_format — "bullets", "report", or "json"
5. Optional: initial source URLs or file_ids to seed the pipeline
6. Handler names (if non-default)

Step 2 — Launch the pipeline

Call workflow_execute:

{
  "workflowIdentifier": "<workflow name or ID>",
  "input": {
    "topic": "<research question>",
    "depth": "medium",
    "output_format": "report",
    "sources": ["<url_or_file_id>"]
  }
}

Note structuredContent.execution_id. Confirm: "Research pipeline started — execution ID: <execution_id>."

Step 3 — Poll and query progress

Loop:

1. Call workflow_status with { "executionId": "<execution_id>" }
2. Check structuredContent.status:
   • COMPLETED → go to Step 5
   • FAILED / TIMED_OUT / CANCELED → surface error and stop
   • RUNNING → continue

While RUNNING, query progress every ~20 seconds:

{
  "executionId": "<execution_id>",
  "action": "query",
  "name": "get_progress"
}

Show the user what the pipeline is doing:

🔍  RESEARCH IN PROGRESS
─────────────────────────
Phase: [phase from result, e.g. "source discovery", "synthesis"]
Sources processed: [N from result]
Hypotheses formed: [N from result]

Also probe for hypothesis checkpoints:

{
  "executionId": "<execution_id>",
  "action": "query",
  "name": "get_hypotheses"
}

If the result contains pending hypotheses awaiting validation, proceed to Step 4.

Step 4 — Handle hypothesis checkpoints

When hypotheses are ready for review (detected from the get_hypotheses query result):

1. Present hypotheses to the user:

   ⏸  HYPOTHESIS CHECKPOINT
   ──────────────────────────
   H1: [hypothesis text]
      Supporting sources: [N]
      Confidence: [low/medium/high]
   
   H2: [hypothesis text]
      ...
   
   Options:
   (A) Validate — continue with these hypotheses
   (B) Amend — provide corrections or constraints
   (C) Inject sources — add documents/URLs to refine
   (D) Discard — restart this hypothesis phase
   

2. Based on user choice, signal the workflow:

   Validate (proceed as-is):

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "hypothesis_decision",
     "input": { "decision": "validate" }
   }
   

   Amend (with corrections):

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "hypothesis_decision",
     "input": {
       "decision": "amend",
       "amendments": "<user corrections or constraints in plain text>"
     }
   }
   

   Inject sources:

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "hypothesis_decision",
     "input": {
       "decision": "validate",
       "additional_sources": ["<url_or_file_id>", "..."]
     }
   }
   

   Discard and restart (hypothesis phase only):

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "hypothesis_decision",
     "input": { "decision": "discard", "reason": "<user feedback>" }
   }
   

Return to Step 3. Deep pipelines may have multiple hypothesis gates.

Step 5 — Deliver the research output

When status === "COMPLETED", present structuredContent.result in the requested format:

✅  RESEARCH COMPLETE
──────────────────────
Topic:       [topic]
Depth:       [depth]
Execution:   <execution_id>
Sources used: [N from result]

[formatted output — bullets / report / JSON as requested]

KEY FINDINGS
────────────
[top 3–5 findings with source citations from result]

CONFIDENCE ASSESSMENT
──────────────────────
[per-finding confidence level + supporting source count]

Offer to pass the output to mistral_chat with magistral-medium-latest + reasoning_effort: "high" for a critical peer review of the findings.