Context from your documents, precision from your database — in one answer.
A Microsoft Copilot Studio assistant that reads each question, decides whether it needs descriptive knowledge or hard data, and loops through both your documents and your database until the answer is complete.
Most assistants only know half the answer.
Single-source assistants force a choice: explain from documents, or look up facts in a database — never both. But real questions mix the two. "Which product suits a family home, and is it in stock?" A document-only bot can't see inventory; a database-only bot can't recommend. The customer gets half an answer.
One topic, two sources, one complete answer.
A Copilot Studio topic reads each question, decides whether it needs descriptive knowledge or precise data, queries the right source, and appends the result to a single running answer — looping until every part of the question is covered. Off-topic questions fall back to web search, so nothing goes unanswered.
A small reasoning engine over two data sources.
Capture & analyze
The user's question is stored and passed to a query-analysis prompt that breaks it into atomic steps and decides which source to query first.
Route
A condition sends each atomic question to the right place — the generative knowledge source or the structured database — depending on what that step actually needs.
Generative source
Descriptions, recommendations, and comparisons come from a PDF attached to the agent as knowledge. If a question falls outside it, web search fills the gap.
Database source
Technical specs, pricing, real-time stock, and dimensions come from a structured Excel table queried as a tool — exact figures, every time.
Compile & check
Each response is appended to one running answer, then a response-analysis prompt judges whether it's enough to fully answer the original question.
Loop or finish
If something's missing, the topic picks the next source and asks again — looping as many rounds as needed until it can deliver one confident, complete answer.
The assistant doesn't pick a source once and hope — it reasons step by step, splitting a request into atomic queries and revisiting either source as many times as needed. That's what lets it answer "recommend something that's actually in stock" in a single turn.
Built entirely on the Microsoft platform.
The whole build lives in flospect.
Every action, variable, prompt, and connection is cataloged — and the documentation is AI-generated from the real flow structure, so it never drifts from what's actually running.
- Focus & scope, foundation summary
- Canvas screenshots embedded inline
- Step-by-step process breakdown
- KPI, monitoring & document-control sections
- Owner, version, and last-modified tracking
- Shareable via a tenant-safe link
Want this running in your tenant?
This is the kind of build we ship in a 4-week Flow Build Sprint — architected, tested, and handed over fully documented in flospect.
Book a discovery call