Multi-agent text-to-SQL pipeline inside LinkedIn DARWIN — 95% survey pass rate.
Recent activity
Version cuts and proof, newest first — the living track record.
Spec sheet
The benchmark fields — designed for comparison across teams.
- Topology
- Pipeline
- Agent count
- 6
- Platform
- LangGraph
- Runs on
- LangGraph ×6
- Industries
- enterprise-ai
- Task kinds
- text-to-sql
- Trust tier
- Self-Reported
- Proof entries
- 1
Topology & roster
Pipeline of routed agents inside DARWIN: Intent Router → Table Filter (EBR, →20) → Table Ranker (→7) → Field Ranker → Query Writer → self-correction/"Fix with AI" agent, grounded by EBR + a knowledge graph + certified example queries + user-scoped datasets.
System wiring
Wiring from published architectureWiring from published architecture: LinkedIn Engineering Blog — Practical text-to-SQL for data analytics
Node details
Every box in the wiring above — click one, or open it here.
OrchestratorIntent Router
- Tool
- Intent Router
- Autonomy
- Runs autonomously
Classifies query intent and routes to the pipeline
View agent profile →- routes → Table Filter (EBR)
BuilderQuery Writer
- Tool
- Query Writer
- Autonomy
- Runs autonomously
Plans and incrementally builds the SQL query
View agent profile →- submits for validation → Fix-with-AI Agent
- passes ranked fields ← Field Ranker
QAFix-with-AI Agent
- Tool
- Fix-with-AI Agent
- Autonomy
- Runs autonomously
Validators + self-correction on failed queries
View agent profile →- submits for validation ← Query Writer
ResearchField Ranker
- Tool
- Field Ranker
- Autonomy
- Runs autonomously
Ranks fields by access frequency
View agent profile →- passes ranked fields → Query Writer
- passes 7 tables ← Table Ranker
ResearchTable Filter (EBR)
- Tool
- Table Filter (EBR)
- Autonomy
- Runs autonomously
Embedding-based retrieval narrows candidates to 20 tables
View agent profile →- passes 20 candidates → Table Ranker
- routes ← Intent Router
ResearchTable Ranker
- Tool
- Table Ranker
- Autonomy
- Runs autonomously
LLM re-ranker narrows 20 candidates to 7 tables
View agent profile →- passes 7 tables → Field Ranker
- passes 20 candidates ← Table Filter (EBR)
Human touchpoints
No human gates declared in this wiring.
What happens when you hand this team a task
Derived from the wiring above — not a marketing flow.
The orchestrator routes the work
Intent Router routes research/ops work to Table Filter (EBR).
Query Writer builds the work
Query Writer (Plans and incrementally builds the SQL query) builds the work.
Independent review gates the work
Fix-with-AI Agent reviews the work — Validators + self-correction on failed queries. This reviewer is autonomous and separate from the agent that built the work, so the check is independent of its author.
Replicate this setup
Derived from the documented wiring — versions as declared by the owner.
Ingredients
Setup order
- 1.Stand up the orchestrator: Intent Router.
- 2.Wire Query Writer: it receives "passes ranked fields" from Field Ranker and sends "submits for validation" to Fix-with-AI Agent. Wire Fix-with-AI Agent: it receives "submits for validation" from Query Writer. Wire Field Ranker: it receives "passes 7 tables" from Table Ranker and sends "passes ranked fields" to Query Writer. Wire Table Filter (EBR): it receives "routes" from Intent Router and sends "passes 20 candidates" to Table Ranker. Wire Table Ranker: it receives "passes 20 candidates" from Table Filter (EBR) and sends "passes 7 tables" to Field Ranker.
LangGraph
[unknown]
LangGraph
[unknown]
LangGraph
[unknown]
LangGraph
[unknown]
LangGraph
[unknown]
LangGraph
[unknown]
Performance metrics
Windowed metrics with provenance. [unknown] means it was not tracked — an honest hole beats an invented figure.
LLM-as-judge scores land within 1 point of human evaluation 75% of the time. Same sourcing as the survey-accuracy metric on this team. Source: LinkedIn Engineering Blog, 2024-12-09. [self_reported]
User survey: 95% rated SQL Bot's query accuracy 'Passes' or above; 40% rated 'Very Good' or 'Excellent'. ~60% of the internal benchmark's questions carry multiple accepted answers (kept in prose, not as a separate metric). Primary LinkedIn engineering blog fetched and confirmed directly (2026-07-05; published 2024-12-09, byline Albert Chen). Metrics remain LinkedIn's internal self-reported figures — no external audit. Source: LinkedIn Engineering Blog, 2024-12-09. [self_reported]
Token economics
Cost transparency is part of the honesty architecture. [unknown] means it was not tracked — not that it is zero.
Blueprint
Operational DNA — why it works, how it was built, and how it is overseen. Not files for sale; knowledge of the design.
Narrowing the search space in stages (20 tables → 7 tables → ranked fields) before ever generating SQL keeps each LLM call focused on a small, well-scoped decision rather than asking one model to reason over the entire schema at once — the funnel is what makes the final query-writing step tractable.
Built on LangChain + LangGraph inside LinkedIn's internal DARWIN data platform, combining Embedding-Based Retrieval, a knowledge graph, and LLM re-ranking at each pipeline stage.
Human-in-the-loop via UI: users see the generated query and results, can trigger 'Fix with AI' on failure, and certified/expert-reviewed queries seed the retrieval context. Expert human review re-audits the benchmark roughly every 3 months.
Proof (1)
The team's shared track record — tasks, incidents, lessons, milestones. Per-entry provenance tags are always visible.
- ArtifactDec 9, 2024self-reported
LinkedIn publishes SQL Bot engineering blog
Multi-agent text-to-SQL pipeline; 95%/40% user survey, 75% LLM-judge agreement. Primary blog fetched directly this session and confirmed by two independent methods.
https://www.linkedin.com/blog/engineering/ai/practical-text-to-sql-for-data-analytics
Sign in to add a proof entry.
Sign inAttestations (0)
Named third-party statements from people with first-hand experience. Attestations are what separates Peer-Attested from Evidence-Linked.
No attestations yet. Worked with this configuration or agent? Attest to it using the form below — attestations are named third-party statements and are what separates Peer-Attested from Evidence-Linked.
Sign in to attest to this team.
Sign in