Every API call is backed by a multi-layer machine learning stack. Here's exactly what runs and when.
All ingestion APIs
Gemini text-embedding-004
Every word and chunk is embedded into 768-dimensional semantic space using Google's Gemini embedding model, with DB caching so repeated words don't re-call the API.
768-d vectors
DB cache
Batch calls
strategy: "rules" (Tier 1)
Linear Projection + Category Signal
768-d Gemini vectors are projected to 256-d via a seeded random orthonormal matrix (preserves cosine distances). A 30% category one-hot signal is blended in to pull same-category words together before MLP training. L2-normalised to unit sphere. Fully deterministic and auditable.
768→256-d
30/70 blend
L2 normalise
QR orthonormal
strategy: "mlp" (Tier 2)
Personal PyTorch MLP
A 3-layer MLP (768→512→512→256) is trained on your word rules using InfoNCE contrastive loss (τ=0.20) plus anchor repulsion (weight=0.35). This learns the decision boundaries between your categories and generalises them to vocabulary not in your rule set. AdamW optimiser, CosineAnnealingLR schedule, gradient clipping at 1.0.
PyTorch
InfoNCE loss
Anchor repulsion
AdamW
Cosine LR
After every ingestion job
Louvain Community Detection
A word co-occurrence graph is built from ingested chunks, then Louvain modularity optimisation groups words into communities. Each community centroid is computed, stored as a pgvector VECTOR(256), and labelled in parallel using Gemini LLM calls.
python-louvain
NetworkX
Parallel LLM labelling
POST /api/prismrag/search
Graph RAG Retrieval Pipeline
Three-phase retrieval: (1) HNSW cosine search on community centroids to find the closest community to the query. (2) BFS 2-hop expansion on the word graph to find related words. (3) MLP re-ranking: if a Tier-2 model exists, all candidates are re-projected through the personal MLP before final cosine ranking.
HNSW m=16
BFS 2-hop
MLP re-rank
pgvector
POST /api/prismrag/bridges
Bridge Vector Synthesis
Cross-community semantic bridges are created by computing the normalised midpoint between two community centroids: normalize((centroid_A + centroid_B) / 2). The bridge vector is then projected through the personal MLP (if Tier 2), creating a synthetic node that connects two previously separate semantic clusters without retraining.
Centroid midpoint
MLP projection
No retraining
POST /api/deliberation/sessions
Multi-Agent Deliberation (LLM ensemble)
The Deliberation pipeline runs a structured ensemble of Gemini LLM calls: one horizontal domain discovery call, up to 10 parallel vertical expert calls (each with domain-specific prompt engineering), and one Master Deliberator synthesis call. Based on the Mixture of Agents (MoA) technique — diverse independent reasoning outperforms a single call on complex questions.
LLM ensemble
Parallel inference
Conflict detection
MoA technique
Every API endpoint
Three-Layer Metering
Usage is enforced and tracked at three levels: Redis sliding window for per-minute rate limiting (<1ms), Redis monthly counter for quota tracking (Postgres fallback), and async Postgres writes for usage_event audit log. Zero DB latency added to the request path.
Redis sliding window
Monthly quota
Async audit log
