Intelligent Business Assistant Web App
Budget: $3,000 – $5,000 USD
What we’re building: A ChatGPT-style web app where company users authenticate (SSO/OAuth) and can chat. The assistant answers questions grounded in company data and always returns citations. Responses stream token-by-token for speed and transparency. Users can thumbs up/down and attach files (PDF/DOCX) for on-the-fly analysis.
How it works (at a glance): 1. User asks a question in chat. 2. Agentic Orchestrator plans the step(s) needed (retrieve, summarize, compare, extract). 3. Search MCP (agentic search) runs retrieval over an embedded corpus spanning Gmail, Google Drive, Notion, and HubSpot. 4. The orchestrator assembles the top-K snippets + instructions and calls the LLM to generate a grounded answer with citations. 5. Feedback (thumbs/reasons) and traces (retrieved IDs/scores) are logged for nightly improvement.
Connectors & Webhooks (read-only, MVP):
-> Sources: Google Drive, Gmail, Notion, HubSpot. Two paths: - Initial Backfill: Enumerate all permitted items, extract/normalize/chunk/embed/upsert with ACL metadata. - Ongoing Updates: Webhooks trigger deltas → re-chunk/embed/upsert. Polling fallback. De-dupe by (source, external_id, etag/version). Agentic Retrieval with Search MCP Input: user query + conversation context. Steps: 1. Embed query (OpenAI embeddings) and call Search MCP. 2. (Optional) Rerank candidate set (MMR or cross-encoder). 3. Assemble prompt (instructions + citations + snippets). 4. Generate with LLM (temperature ≤ 0.4, “prefer provided context, say ‘I don’t know’ if insufficient”). 5. Return streaming answer + structured citations.
Data freshness & reliability targets:
-> Freshness: Drive/Notion ≤ 5 min · Gmail ≤ 10 min · HubSpot ≤ 5 min. -
MVP deliverables (what “done” looks like):
-> Chat UI with streaming + AUTH , citations, feedback. - Search MCP backed by supabase/pgvector over Gmail/Drive/Notion/HubSpot content. - Connectors supporting backfill + webhook/polling deltas that auto-embed and upsert.
How it works (at a glance): 1. User asks a question in chat. 2. Agentic Orchestrator plans the step(s) needed (retrieve, summarize, compare, extract). 3. Search MCP (agentic search) runs retrieval over an embedded corpus spanning Gmail, Google Drive, Notion, and HubSpot. 4. The orchestrator assembles the top-K snippets + instructions and calls the LLM to generate a grounded answer with citations. 5. Feedback (thumbs/reasons) and traces (retrieved IDs/scores) are logged for nightly improvement.
Connectors & Webhooks (read-only, MVP):
-> Sources: Google Drive, Gmail, Notion, HubSpot. Two paths: - Initial Backfill: Enumerate all permitted items, extract/normalize/chunk/embed/upsert with ACL metadata. - Ongoing Updates: Webhooks trigger deltas → re-chunk/embed/upsert. Polling fallback. De-dupe by (source, external_id, etag/version). Agentic Retrieval with Search MCP Input: user query + conversation context. Steps: 1. Embed query (OpenAI embeddings) and call Search MCP. 2. (Optional) Rerank candidate set (MMR or cross-encoder). 3. Assemble prompt (instructions + citations + snippets). 4. Generate with LLM (temperature ≤ 0.4, “prefer provided context, say ‘I don’t know’ if insufficient”). 5. Return streaming answer + structured citations.
Data freshness & reliability targets:
-> Freshness: Drive/Notion ≤ 5 min · Gmail ≤ 10 min · HubSpot ≤ 5 min. -
MVP deliverables (what “done” looks like):
-> Chat UI with streaming + AUTH , citations, feedback. - Search MCP backed by supabase/pgvector over Gmail/Drive/Notion/HubSpot content. - Connectors supporting backfill + webhook/polling deltas that auto-embed and upsert.