Self-Hosted Deep Research over Customer Comms (L5)
Self-hosted vector DB of all customer calls/emails/surveys with a hybrid (semantic + text) re-ranker. An agent harness fans out: project → queue → per-deal analyzer → synthesizer → HTML report. The reason most AI search tools feel wrong: they pretend to read every deal but actually rank-and-snip. From Egan at Sendoso on the GTM AI Podcast.
The steps
- 01
Own your customer-comms vector store
Tool: Self-hosted
Build your own vector DB seeded with call transcripts, emails, survey responses. "No matter what provider we use, we're always going to have our own set of data that AI can reliably talk to." Provider switching cost drops to zero.
- 02
Use a hybrid re-ranker (semantic + text)
Tool: Edge function
Semantic search alone misses exact terms. Text alone misses meaning. Run both, score each, re-rank by combined score. This is where chunking strategy and embedding choice actually matter.
- 03
Use the Anthropic agent-harness pattern
Tool: Anthropic agent harness
Don't stuff one agent with 200 deals, it will lie. Pattern: orchestrator creates a project, queues every deal, fires a per-deal analyzer (full context for THAT deal), then a synthesizer rolls all results up. Same pattern Claude Code uses.
- 04
Output as a single-file HTML report
Tool: Claude / GPT
Skip slides. Generate a portable single-file HTML doc per request, win/loss, stage-2 objection patterns, churn drivers. Host internally behind auth.
- 05
Treat hallucination as a context problem
Tool: Prompt engineering
Egan: "A lot of hallucination isn't hallucination, it's missing context." Aggressively prompt-engineer in your product terms, contract vocab, internal acronyms. Prompting matters MORE in full agentic builds, not less.
Tools in this playbook
- Self-hosted
- Edge function
- Anthropic agent harness
- Claude / GPT
- Prompt engineering
Next playbooks
Unfamiliar terms are defined in the AI and Revenue Dictionary. Related frameworks live in the framework library.