Flow Retriever (Virtual Data Source)

​

How It Works

User sends a message
  └── LangChain Orchestrator
       ├── Azure AI Search (index-based sources) ────── documents
       ├── Flow Retriever (flow-based sources) ───────── documents
       ├── Vector Retriever (external vector stores) ─── documents   ──→  merged into RAG context
       └── Web Search (if enabled) ───────────────────── documents

1. The orchestrator identifies FLOW_RETRIEVER entries in the chat’s dataSources[]
2. Each retriever source is executed in parallel via performFlowRetrieverRAG()
3. The flow receives { query, connectionId, _retrieverInstructions } as inputs
4. The flow’s final output is parsed as structured documents
5. Documents are merged into the RAG context for the LLM to cite

​

Execution Contract

• JSON array — [{ "title": "...", "content": "...", "url": "..." }]
• JSON object with documents key — { "documents": [...] } or { "results": [...] }
• Plain text — Wrapped as a single document automatically

​

Connection Modes

​

Configuring a Chat with Flow Retriever

1. Open a chat’s Edit Form → Data Sources
2. Check Flow Retriever
3. Select a Retriever Flow from the dropdown (only flows with flowType: retriever appear)
4. Choose a Connection Mode:
   • Fixed — pick a connection from the dropdown
   • User selects at runtime — users choose in the sidebar
5. Optionally add Retriever Instructions — natural-language guidance injected as {{_retrieverInstructions}}
6. Save the chat

​

Creating a Retriever Flow

1. Set flowType to retriever — In the flow editor, set the flow type to “Retriever”. This marks the flow as headless and makes it available in the Flow Retriever dropdown.
2. No user interaction nodes — Retriever flows must not contain FORM_PROMPT or HUMAN nodes. They execute headlessly with no user interaction at execution time.
3. Accept the standard inputs — The Start node should expect query (the user’s search text) and optionally connectionId and _retrieverInstructions.
4. Return structured documents — The final LLM/output node should produce a JSON array of documents:
   [
     { "title": "Row 1", "content": "The data from this row...", "url": "optional-link" },
     { "title": "Row 2", "content": "Another result..." }
   ]
   

​

Example: Database Retriever Flow

Start → LLM (NL→SQL) → Tool (Execute SQL) → LLM (Format Results) → End

You are a SQL query generator. Convert the user's natural-language question
into a valid SQL query for the connected database.

{{_retrieverInstructions}}

User question: {{query}}

Respond with ONLY the SQL query, no explanation.

Format the following SQL query results as a JSON array.
Each element must have "title" and "content" fields.
The title should be a short identifier. The content should be
a natural-language summary of the row data.

Results:
{{sql_results}}

Respond with ONLY the JSON array.

​

Bootstrapped Retriever Flows

​

Security

• Connection ACL — Every connection is checked against the user’s entitlements before execution
• Flow validation — Only flows with flowType: retriever are accepted
• User identity injection — _userId, _userUpn, _userEmail are injected server-side for row-level filtering
• Timeout — Each flow execution has a 30-second timeout boundary
• Parallel execution — Multiple retriever sources execute in parallel; failures are isolated per-source

​

IDataSourceConfig Fields (Flow Retriever)

See Datasource Catalog → IDataSourceConfig Fields for the full field reference. Flow Retriever–specific fields are summarized below.