SecondActSociety — RAG Pipeline V2 Implementation Planning

v2 | 5/5 Decisions Confirmed | 2026-03-07
From: Jonathan (noboxAI)  |  To: Skip
Architecture Validated — Implementation Ready
All 5 Architecture Decisions Confirmed

Every decision from the v1 briefing was confirmed by Skip. The architecture is validated and ready for implementation sprint planning.

Latency Target
Connection Pool Warming
Observability Stack
LangSmith Only
Enhancement Priority
Conversation Memory (LIVE)
Graph Traversal Depth
3 Chunks
Fallback Strategy
Broad Match
Conversation Memory is LIVE
Spec Correction

The v1 briefing listed conversation memory as a deferred enhancement priority. Skip confirms this is already implemented — the 5 most recent messages/turns are included in retrieval context via Redis session state. This corrects the v1 specification and removes conversation memory from the enhancement backlog.

Updated Pipeline — Conversation Memory Active

    

Redis session persistence is active. Each query includes the 5 most recent conversational turns, enabling context-aware retrieval and response generation. The session state informs intent classification confidence and provides continuity across multi-turn interactions.

Confirmed Implementation Priorities

Latency: Connection Pool Warming

What: Pre-warm database connections on service startup plus periodic keepalive to maintain warm connection pools.

Why: Railway's ephemeral containers can go idle. When a container has been idle, the first user request incurs a cold-start penalty as database connections are re-established. Connection pool warming is the confirmed #1 latency optimization target.

How: FastAPI startup event initializes connection pools for PostgreSQL/pgvector, Neo4j, and Redis. A background keepalive task periodically pings each connection to prevent idle timeout.

Observability: LangSmith (Current Stack)

What: LangSmith tracing confirmed as sufficient for current observability needs.

Config: LANGSMITH_TRACING=true, project second-act-learning-v2

No additional tooling required at current scale. LangSmith provides trace-level visibility into LangGraph node execution, retrieval quality, and response generation latency.

Confirmed Parameters (No Changes)

All retrieval parameters are locked at their current values following v1 confirmation. No tuning required for implementation sprint.

Parameter Confirmed Value Status
PERSONA_RETRIEVAL_COUNT 3 Locked
FALLBACK_THRESHOLD 3 Locked
VECTOR_WEIGHT 0.5 Locked
RRF_K 60 Locked
MAX_RAG_RESULTS 12 Locked
Next Enhancement Priority

With conversation memory shipped, the next enhancement slot is open. The following candidates are ranked by estimated impact.

Enhancement Candidates — Weighted Priority

    
Your Input

With conversation memory shipped, what's the next enhancement priority?

  • B) A/B Testing Framework — Experiment with retrieval parameters (RRF weights, phase ratios, fallback thresholds). Requires sufficient query volume for statistical significance.
  • C) Authority Taxonomy Evolution — Add new authority levels or reclassify sources as content library grows. Currently: primary, supplementary, illustrative, supporting.
  • D) Multi-Modal Content Ingestion — Add video/image content paths to the RAG pipeline. Activates when Jay produces visual content materials.
Connection Pool Warming Scope
Your Input

For connection pool warming, which implementation scope?

  • A) Service Startup Only — Pre-warm all connections (PostgreSQL/pgvector, Neo4j, Redis) on FastAPI startup event. Simple and reliable.
  • C) Lazy + Circuit Breaker — No pre-warming. Instead, add circuit breaker pattern for graceful degradation when connections are cold.
Implementation Timeline

Sprint plan based on confirmed priorities. The next enhancement slot depends on v2 input.

Implementation Sprint Plan