Companion App

From Pipeline to Product

Modules 1-4 built the engine — data ingestion, encoding, retrieval, and the AI gateway. Now we wrap it in a product that sales reps actually want to use. The Companion App is a chat interface with one killer feature: pre-call briefings that pull together everything the system knows about an account.

Key Concepts

Streaming Responses

Users expect AI to start responding immediately, not wait 10 seconds for a complete answer. Streaming sends tokens as they're generated:

User: "Brief me on Acme Corp"

[200ms] "Acme Corp is a "
[250ms] "mid-market SaaS company "
[300ms] "currently in the negotiation stage..."

Technically, this is Server-Sent Events (SSE) — the server pushes chunks to the client over a single HTTP connection. The client renders each chunk as it arrives, creating the typewriter effect.

The Briefing Agent (LangGraph.js)

A pre-call briefing needs data from multiple sources — CRM, transcripts, tickets, competitors — and each search is independent. Running them sequentially is slow. LangGraph.js enables a fan-out pattern:

                   ┌──→ Fetch CRM Data ──────┐
                   │                          │
Parse Request ──→  ├──→ Fetch Call History ───┤──→ Synthesize ──→ Format
                   │                          │
                   ├──→ Fetch Support Tickets ┤
                   │                          │
                   └──→ Fetch Competitor Intel ┘

All four fetches run in parallel. The synthesize node waits for all of them, then combines the results into a coherent briefing. What would take 8 seconds sequentially takes 2 seconds with fan-out.

Graph Composition

The briefing agent graph composes with the gateway graph from Module 4. When a user asks for a briefing:

This is the power of graph-based orchestration — complex workflows compose without becoming spaghetti code.

Source Citations

Enterprise users don't trust AI that says "trust me." Every claim in a response must cite its source:

Acme's main concern is the integration timeline [1].
Their VP of Engineering, Jane Smith, mentioned needing
completion before Q2 board meeting [2]. There's also an
open support ticket about API latency [3].

Sources:
[1] CRM - Acme Corp (updated 2024-03-15)
[2] Call Transcript - Acme Q4 Review (2024-02-28)
[3] Support Ticket #1847 (2024-03-01)

Implementation: the context assembly step (Module 3) tags each chunk with a source reference. The AI prompt instructs the model to cite these references. Post-processing maps citation numbers to full source details.

Confidence Scores

Not all AI responses are equally reliable. Confidence scoring signals to the user when to trust and when to verify:

This is computed by analyzing: how many source chunks were used, how similar they were to the query, how recent the data is, and whether multiple sources corroborate the answer.

Graph Execution Trace

For debugging and transparency, every graph execution produces a trace:

{
  "trace_id": "brief-acme-001",
  "nodes_executed": ["parse", "fetch_crm", "fetch_calls", "fetch_tickets", "fetch_competitors", "synthesize", "format"],
  "duration_ms": 2340,
  "tokens_used": { "input": 3200, "output": 890 },
  "sources_retrieved": 12,
  "sources_cited": 3
}

This trace is invaluable for debugging ("why did it miss that ticket?"), optimization ("which fetch is slowest?"), and auditing ("what data did it access?").

What You'll Build

Glossary