SecondAct Ingestion Pipeline

Parallel Streams Architecture
FROM: Jonathan Gudger, Technical Architect TO: Jay Samit March 4, 2026
8 decisions awaiting Jay's input

From 2 Sources to 5+ Parallel Streams

Digital Jay's knowledge currently comes from 2 sources (Brain + Voice). The ingestion pipeline expands this to 5+ parallel streams, each with distinct content types, owners, quality gates, and ingestion cadences.

The Problem

Jay tested Digital Jay at a senior community tech hall — quality is high but scope is narrow. Specific gaps identified:

Gap

"Who does Jay admire?"

Can only pull names from the manuscript. Needs personal opinions and preferences.

Gap

"My wife just died"

Handled via book's friendship themes, but needs dedicated grief/loss content with sensitivity guardrails.

Gap

"What new AI tools came out?"

Zero knowledge of current events. Jay's #1 missing feature request.

Gap

Personality depth

Book Jay vs. Real Jay. Needs informal personality, humor, personal stories.

Parallel Streams Architecture

The ingestion pipeline processes 5 distinct content streams through a unified vector store + Neo4j graph, with source-aware routing at query time.

Full Pipeline — 5 Parallel Ingestion Streams

Current vs. Target State

Where we are today, where this pipeline gets us, and the full ecosystem vision.

2-Source Current State to 5+ Source Target

5 Content Streams, Each Purpose-Built

Each stream has distinct content characteristics, owners, ingestion cadences, and quality requirements.

1

Personal Q&A

Give Digital Jay personality depth beyond the book
Content: 50-70 curated Q&A pairs Owner: Jay writes, Skip reviews Cadence: Initial batch, then drip-feed

Opinions, preferences, heroes, humor, personal history. Sources include "roommate lists," StoryCorps questions, and user feedback gap analysis. Jay types answers — "I type as fast as I talk."

Q&A Ingestion Workflow
2

Life Transitions (Grief/Loss)

Empathetic responses to grief, divorce, identity loss
Content: Modular docs, 500-1,500 words each Owner: Jay writes, Skip sensitivity review Cadence: Review gate before every ingest

Jay's core demographic territory. NOT for public posting — exclusively for RAG ingestion. Requires sensitivity review + crisis resource linking. Every emotional support response includes 988 Lifeline link.

Sensitivity Guardrails (Non-Negotiable)

GuardrailImplementation
Crisis resource links988 Suicide & Crisis Lifeline appended to every emotional support response
Scope disclaimer"I'm not a therapist, but here's what Jay has written about..." prefix
Content review gateTeam reviews ALL grief content before ingestion — no auto-ingest
No diagnostic languageNever suggest diagnosis, never minimize pain
Escalation pathIf user expresses immediate distress, crisis resources prominently displayed
Grief/Life Transitions Pipeline
3

AI News Pipeline

Jay's #1 feature request — "What's going on in AI"
Content: Newsletter digests, 3-4 sources Owner: Jonathan (manual Phase 1) Cadence: Weekly manual, then daily automated

Phased approach: manual prototype (Mar 10-21) with 1 source, team evaluates Mar 21, automate if it moves the needle. Attribution critical — every news response prefixed with "As of [date]..."

PhaseActionTargetGate
0. Source IDJay + Skip identify 3-4 newslettersBy Mar 7Sources identified
1. ManualWeekly manual ingestion of 1 sourceMar 10-21Keynote complete
2. EvaluateTest Digital Jay with news queriesMar 21 callDoes it move needle?
3. AutomateScheduled pipeline: scrape, parse, embedPhase 2 sprintGo decision
News Pipeline — Phased Rollout
4

Resource Pages

Deep-dive guides Digital Jay can reference and link to
Content: Long-form guides, 1,000-3,000 words Owner: Skip drafts, Jonathan publishes Cadence: Manual trigger per page

Each page has email capture CTA + Digital Jay prompt suggestions. In Reader/Advisor and Community modes, Digital Jay appends "Learn more: [URL]" links. First 2 recommended pages: Self-Assessment Questionnaire and BOSS Framework Quick Reference.

Resource Page Flow
5

Voice (Interviews/Podcasts)

Already Live — No New Decisions Needed

Transcribed audio from podcasts, interviews, and media appearances. Already operational — growing corpus as Jay's media schedule expands.

Intent Classification and Source Selection

How the LangGraph agent service decides which streams to query for any given user question.

Knowledge Source Routing Logic

Source Priority Matrix

Query TypePrimary SourceSecondaryTertiarySpecial Rules
Book contentBrainVoice--Standard RAG
Personal questionPersonalVoiceBrainPersonality + context
Emotional supportLife TransitionsBrain--Crisis links + disclaimer
Framework/methodBrainResources--Link to resource page
Current eventsNews----Date prefix + attribution
Career adviceBrainPersonalResourcesMulti-source synthesis
"What does Jay think?"PersonalBrainVoiceOpinion-aware routing

Mode-Aware Response Rules

ModeLengthResource LinksNews AttributionCrisis Resources
Press~150 wordsNeverDate onlyIf triggered
Reader/Advisor~400 wordsInline + footerFull attributionIf triggered
Community~400 wordsInline + footerFull attributionIf triggered

8 Decisions That Unblock the Pipeline

These decisions are blocked on Jay. Each one directly affects how we build, what we build, and what goes live first.

Ready to make your decisions? Open the interactive input form to record your choices.

Open Input Form
REQ-1: Where should the Q&A questions come from? HIGH
Recommended: C — Custom curated from user gaps. Combines roommate lists + StoryCorps + real user query gaps for maximum coverage.
Options: A) "Roommate" lists (~100 Qs) | B) StoryCorps deep questions (~50) | C) Custom curated from team
REQ-2: How should Jay deliver answers? MEDIUM
Recommended: A — Written (typed). Highest quality, no transcription errors, direct ingestion. Jay: "I type as fast as I talk."
Options: A) Written (typed) | B) Voice recorded, transcribed | C) Video, transcribed
REQ-3: Should Q&A answers get their own knowledge source? HIGH
Recommended: B — New "Personal" source. Enables opinion-aware routing — personal questions route to personal content, not book content.
Options: A) Merge into Brain | B) New "Personal" source | C) Merge into Voice
REQ-4: How broad should the grief/loss content scope be? HIGH
Recommended: C — Full emotional transitions (grief, divorce, identity, loss, reinvention). Jay's audience is 40-65 — all of these hit.
Options: A) Grief only | B) + gray divorce | C) Full emotional transitions
REQ-5: What format for grief/loss content? MEDIUM
Recommended: C — Both essays and Q&A. Essays provide depth, Q&A handles specific scenarios. Best coverage.
Options: A) Essay-style | B) Q&A format | C) Both
REQ-6: Approve sensitivity guardrails for grief responses? CRITICAL
Recommended: C — Full guardrails. Crisis links + "not a therapist" disclaimer + escalation path. This is non-negotiable for responsible AI.
Options: A) No special handling | B) Crisis links only | C) Full guardrails (crisis + disclaimer + escalation)
REQ-7: What should the grief knowledge category be named? MEDIUM
Recommended: B — "Life Transitions" source. Broad enough to include grief, divorce, and identity — doesn't pigeonhole the content.
Options: A) Fold into existing book chapters | B) New "Life Transitions" source | C) New "Emotional Support" source
REQ-8: Which AI news sources should we use? HIGH
Recommended: D — Jay + Skip identify 3-4 trusted sources. You know what you trust; we'll validate RSS/API feasibility.
Options: A) TLDR | B) The Neuron | C) Ben's Bites | D) Jay + Skip select

8 decisions, 5 minutes. Your answers unblock 3 engineering streams and 2 content workstreams.

Open Input Form

Dependency Map and Gate Checklist

Current streams validate the ingestion path. Once proven, deferred streams (automated news, community, courses) activate through the same architecture.

Stream Dependencies and Activation Gates

Gate Checklist

  • Brain (Book) pipeline — Live. 1,403 nodes from 318-page manuscript.
  • Voice (Interview) pipeline — Live and growing. Bifurcated RAG operational.
  • Personal Q&A pipeline — Blocked on: Jay's decision on question source (REQ-1) + delivery format (REQ-2).
  • Life Transitions pipeline — Blocked on: Grief scope (REQ-4) + guardrails approval (REQ-6).
  • Resource Pages pipeline — Blocked on: Skip's first 2 resource page drafts.
  • News pipeline (manual) — Gate: Keynote complete (Mar 6-7), then manual prototype Mar 10-21.
  • News pipeline (automated) — Gate: Manual prototype evaluation at Mar 21 call.
  • Community content — Gate: Community platform launches post-book launch (April 2026).
  • Course content — Gate: Kajabi course digitized and structured.

Key Quotes from the Meeting

Direct quotes from the Feb 23 Skip/Jay weekly check-in that shaped this pipeline architecture.

"I've yet to have an answer where I go, holy fuck. But it's answering with limited stuff." — Jay Samit, on Digital Jay quality vs. scope
"That's the single most important missing thing." — Jay Samit, on AI news awareness for Digital Jay
"I type as fast as I talk, and I edit." — Jay Samit, on delivery format for Q&A answers
"My wife just died" — this is the kind of query we need to handle with real care. Jay's audience is 40-65; grief, divorce, and identity loss are realities, not edge cases. — Team discussion, on Life Transitions content need