Sprocket document pipeline — tool selection

Working note for the KwikTrip AI infrastructure decision. Founder asked for UiPath background, then for the Databricks equivalent and how it integrates with Copilot Studio.

The problem being solved

Sprocket is a conversational support agent for the store engineering team, surfaced through a Teams DM. Equipment fails at a service location — hot dog roller, fuel pump, car vacuum — a work order is issued, the tech photographs the serial number and describes the fault, and the agent returns instructions drawn from a manual knowledge base.

Currently ~80% accuracy on helpful recommendations. The bottleneck is retrieval from the manuals, especially diagrams. A document pre-processing pipeline is being built to improve context retrieval.

Stack context: Copilot Studio (agent + Teams distribution), Databricks (licensed), UiPath (present internally). Snowflake has been scrapped.

Why diagrams break RAG

Worth stating plainly because it determines tool choice: a diagram is pixels. Text-only chunking gives the retriever nothing to match a query against, so the figure that answers the question is invisible at retrieval time even when it sits in the indexed document. The fix is to give diagrams text — descriptions — and to keep them anchored to the procedure text around them.

UiPath — background and honest fit

Founded 2005 in Bucharest as DeskOver (Daniel Dines, Marius Tîrcă); automation services first, productized RPA ~2012, renamed UiPath 2015. NYSE IPO April 2021 (PATH), among the largest US software IPOs at the time. LLMs undercut the core RPA premise; Dines returned as CEO May 2024 to convert an RPA company into an enterprise AI automation platform.

FY2025 revenue $1.43B, +9% YoY; ARR $1.67B. A company of that size growing 9% is defending a category, not winning one. Relevant to a multi-year architecture bet.

Real strengths: orchestration, governance, audit trails, human-in-the-loop approvals, large installed base. Low-code (Studio / StudioX / Studio Web), not no-code, and UiPath for Coding Agents now lets a coding agent build IXP flows — this answers the founder's "can an agent help build it" question directly.

Relevant products: IXP (Intelligent Xtraction and Processing) + Context Grounding (their RAG indexing layer; indexes created in Orchestrator, attached to agents in Studio Web).

The fit caution: IXP is built to turn semi-structured business documents into structured data — invoices, receipts, forms, contracts. Pull the invoice number, get the line items. That is a different engineering problem from retrieving procedural content and diagrams out of technical service manuals. The two look alike on a slide. Do not assume transfer.

Databricks — the Snowflake translation

The load-bearing capability: ai_parse_document parses PDFs and explicitly captures tables, figures and diagrams with AI-generated descriptions and spatial metadata, storing results in Unity Catalog so documents behave like queryable tables. Databricks claims leading quality at 3–5× lower cost, at millions-of-documents scale.

That targets the exact failure mode. Generated descriptions make a diagram matchable; spatial metadata preserves its relationship to surrounding procedure text.

Pipeline, all in-platform: ingest → ai_parse_document → ai_extract / ai_classify (v2 signatures, natively composable with the parse step) → Vector Search index → retrieval.

Integration with Copilot Studio

Databricks ships managed MCP servers exposing vector search endpoints, Genie spaces and UC functions over a native MCP interface (/api/2.0/mcp/...). Copilot Studio adds an MCP server directly as a Tool. The agent stays in Teams, retrieval executes in Databricks, no custom glue.

Two details that bite late if nobody raises them:

1. Licensing. MCP server added directly as a tool in Copilot Studio → no additional licensing beyond standard Copilot Studio. The same integration built as a Power Apps custom connector → Premium licensing for every end user. Architecturally near-identical, materially different per-head cost across a store engineering team. Go MCP.
2. Authentication. End-user auth preserves Unity Catalog RBAC and fine-grained access control. The maker-credentials alternative works and silently bypasses user-level access control — acceptable for a demo, not for a client system whose governance story is Unity Catalog.

Front-end / deployment paths (both platforms have one)

Databricks — three tiers, no Copilot Studio required:

• Genie One (GA 2026-06-16) — the direct Snowflake Cowork / Snowflake Intelligence equivalent. Full-screen natural-language interface for business users; searches Genie Agents first, then dashboards, queries and metric views. Chats on mobile, Slack, Teams, or any productivity tool.
• Databricks Apps — secure customizable chat interface, serverless compute, built-in SSO. The custom front-end path.
• Genie App Builder — generates a business app from a natural-language description, grounded in Lakehouse data, governed by Unity Catalog.
• Databricks One — curated business-user "front door."
• Agent Bricks now supports Claude Code SDK, LangGraph, Agno, CrewAI and OpenAI Agent SDK, with horizontal autoscaling via Databricks Apps.

So Databricks could deliver Sprocket end to end, Teams included. Which raises the unasked question: is Copilot Studio load-bearing here, or just where the project started?

The caveat that may decide it — input modality. Sprocket's workflow begins with a technician photographing a serial number in the field. Copilot Studio in Teams handles that upload path natively. Whether Genie One treats image input as first-class is unverified — its documented strength is natural-language questions over data. This is likely a bigger determinant than any parsing benchmark, and it is cheap to check directly.

UiPath — correction to the earlier "third platform" framing. That undersold it:

• Bi-directional Copilot Studio integration shipped with Microsoft. The enhanced Power Platform connector embeds UiPath agents and automations into Copilot Studio as actions, topics and agent flows; Copilot agents can conversely be embedded in UiPath Studio.
• UiPath Maestro orchestrates Copilot Studio agents alongside UiPath agents, robots and people.
• Enhanced Autopilot agent for Copilot for M365 and Teams — so UiPath has its own native front-line surface, and it also reaches Teams.
• MCP integration with Azure / AI Foundry.

This is a genuine Microsoft partnership, not a bolt-on. The integration objection was weaker than originally stated and should not carry weight in the decision.

Layer mapping — avoid a muddled bake-off

Parsing and retrieval are different layers and were being compared across each other:

Layers can be mixed — parse on one platform, retrieve on another — but comparing IXP to Vector Search directly will produce a bad evaluation.

Recommendation

Databricks, not UiPath, for the document pipeline — but on a narrower argument than the first draft of this note made.

The integration objection is withdrawn: UiPath's bi-directional Copilot Studio integration is first-class, and it has its own Teams front-line path. That is not a reason to rule it out.

What remains, and it is sufficient: IXP's demonstrated strength is structured field extraction from semi-structured business documents. ai_parse_document is explicitly built for tables, figures and diagrams with generated descriptions and spatial metadata, at scale and low cost. The bottleneck is diagrams. Databricks is also already licensed, which makes the marginal cost argument agree with the capability argument.

On the front end: keeping the agent in Copilot Studio remains reasonable, but the reason has changed. It is no longer "Teams distribution is unique" — Genie One and UiPath Autopilot both reach Teams. It is that Copilot Studio's photo-upload path suits Sprocket's multimodal field workflow, and that advantage is real only until Genie One's image-input support is checked.

What is NOT established

Applied symmetrically to both vendors, so the bar does not move between them:

• Nothing here is from having built it — docs and community write-ups only.
• Whether ai_parse_document handles a real exploded-parts diagram is a pilot question, not a spec question. Run it against ~20 manual pages that are currently failing retrieval. One day of work, and it settles the debate with evidence instead of vendor claims.
• Technique may matter more than vendor regardless: layout-aware chunking that keeps a figure with its caption and steps, multimodal page-image embeddings rather than text-only, and serial-number → model → manual-section routing before retrieval runs.

Open

• Run the 20-page pilot.
• Confirm whether KwikTrip's UiPath licensing is sunk/enterprise-wide (changes the marginal-cost argument, though not the fit argument).
• Snowflake was scrapped — no bearing on the founder's Snowflake GenAI cert (2026-08-24), which is a market-wide credential, not a client-specific one.