Channels + triggers + schedules
A chat request enters the live streaming pipeline with its task intent.
Case Studies / Enterprise AI Hub
An enterprise prompt library with a contract-review system.
Built for a global healthcare company's internal innovation program under the Maslow AI brand: a department-organized prompt library backed by a Statement-of-Work review system with graph retrieval, field-level citations, and structured answers.
7
generative-UI widgets available for structured responses
PRODUCTION EVIDENCE
93%
expected first-tool routing on a 28-case live test suite
PRODUCTION TEST RESULT
50
SOW corpus, chunked by document structure
PRODUCTION EVIDENCE
11
departments in the curated prompt library
PRODUCTION EVIDENCE
EXECUTIVE SUMMARY
The case in four decisions
- Waiting work
- Knowledge workers lacked a shared prompt library and a fast way to review dense Statements of Work with traceable source fields.
- What changed
- A department prompt library and SOW reviewer now combine vector and graph retrieval, structured response tools, and field-level citations.
- Human decision
- Procurement and legal reviewers inspect cited source fields and decide whether drafted answers or documents can move forward.
- Evidence state
- The deployed implementation and 26-of-28 first-tool routing result are production evidence. Adoption and business-impact measurement are still in progress.
THE CHALLENGE
The product needed to support two jobs
Enterprise knowledge workers had no shared home for reviewed prompts, and no fast, trustworthy way to interrogate dense Statement-of-Work documents. "What are the termination terms?" meant an afternoon of reading. "Where's the duplicate-spend risk across these vendors?" meant nobody checked.
AgentHub combines a curated prompt library, organized by department, with an SOW review system grounded in the source documents.
THE EXPERIENCE
A prompt-library front door with a grounded review engine
THE FRONT DOOR
Department prompt library
Eleven color-coded departments, live search, likes and comments, public/private prompts, per-author categories. Every prompt opens straight into the AI chat; the library is how people discover what the review system can do.
MarketingLegalR&DFinanceSupply Chain+6 more
THE ENGINE
SOW review system
Ask a contract corpus questions in plain English (“compare the day rates across these SOWs”) and get grounded answers with citations to the exact field in the source document, via parallel dense-vector and knowledge-graph retrieval.
▸ VECTOR SEARCH · top-5 · 3072-dim embeddings▸ KNOWLEDGE GRAPH · entities + one-hop neighbors▸ CITATIONS · down to the exact jsonPath field
DEEP DIVE · HYBRID RAG
From 50 contracts to field-level answers
Parallel dense-vector and knowledge-graph retrieval, fused as grounded context. Drafted document sections can cite the exact source field. If the graph is unavailable, retrieval degrades gracefully to vector-only.
EXPLORE THE SIX-STEP RETRIEVAL PIPELINE
01INGEST
Structure-aware chunking
Contracts follow their document structure during chunking, with breadcrumb metadata (section path and field names) preserved on every chunk.
02EMBED
3072-dim embeddings
Each chunk is embedded and stored behind a native vector index: cosine similarity over the full corpus.
03RETRIEVE ×2
Two legs in parallel
Dense vector search (top-5) runs alongside a knowledge-graph query for matching entities (vendors, clients, deliverables) plus their one-hop neighbors.
04FUSE
Grounded context fusion
Both result sets enter the prompt as labelled context sections, each vector hit carrying its similarity score and source path.
05GENERATE
Multi-step response loop
The model answers in a multi-round tool loop, streaming tokens, reasoning and widgets live over server-sent events.
06CITE
Field-level citations
Drafted documents cite per-section back to the exact field of the source SOW, auditable by a procurement or legal reviewer.
EXAMPLE TELEMETRYA data-source badge can show a vector result count and top score or a graph entity count. The Activity Panel displays retrieval searches, tool rounds, and source paths during a review.
DEEP DIVE · INTENT ENGINEERING & GENERATIVE UI
The review system selects what kind of response to give
Every message is routed through a deterministic intent layer: tool request, document drafting, data-heavy, narrative, off-topic or conversational. That classification gates which tools the model is offered. Conversational turns keep chart tools unavailable.
EXPLORE RESPONSE ROUTING AND TOOL EXAMPLES
SEVEN WAYS TO ANSWER
LIVE TEST RESULT
26 / 28
Cases routed to the expected first tool through the live streaming pipeline (92.9%).
“put the comparison in a table” →
show_data_table“draft a PRD” →
ask_questions first, then the document“what's the weather in Boston?” → text only, no tool
REVIEW CONTROLS
Traceable implementation controls for human review
REVIEW CONTROLS IN THIS IMPLEMENTATION
✓Visible retrieval record: the Activity Panel displays retrieval steps, tool calls, and source paths
✓Field-level citations: auditability that procurement and legal reviewers can check
✓Graceful degradation: if the graph is unavailable, the system records a vector-only fallback
✓Tool availability control: function-calling disabled for conversational turns
✓Routing regression coverage: the 28-case intent test suite plus browser test specs, run against the live pipeline
EXPLORE THE IMPLEMENTATION STACK
UNDER THE HOOD
Frontend: React + TypeScript, Tailwind, assistant-ui with a custom streaming runtime, Recharts
Data: Convex reactive DB with native vector index · Neo4j knowledge graph · Clerk auth with post-sign-in intent replay
AI: Gemini 3 Flash (chat, thinking, tool-calling) · gemini-embedding-001 at 3072 dimensions · optional MCP tool discovery at runtime
Delivery: serverless; the whole chat/RAG/tool loop runs as one streaming action
The system provides field-level citations for graph-augmented SOW review and uses intent routing to select response tools. Dense and graph retrieval are combined into one context packet; the intent route uses deterministic orchestration.
HOW IT WAS BUILT · DEPLOYED IMPLEMENTATION
The demonstrated components in shared language
This map is scoped to components demonstrated in the deployed AgentHub implementation. Open the optional detail to compare the responsibilities.
VIEW THE DEPLOYED ARCHITECTURE MAPPING
Context + session and durable memory
Vector and graph retrieval assemble cited contract context.
Skills + orchestration
Deterministic intent routing governs the multi-round tool loop.
Tools + connectors + delegation
The classified request gates which response tools are available.
State + trace + checkpoints
The Activity Panel exposes searches, tool rounds, and source paths.
SERVICES USED
This engagement, as catalog services
MASLOW DELIVERY SUMMARY
The prompt library is the front door. The engineering story is a contract-review system that shows its sources, down to the exact field.
Live demo deployed · 11-department library · 50-SOW corpus · adoption metrics in progress
Have documents your teams are afraid to ask questions of?
We build grounded, citable review systems within the agreed data and access boundary.