How It Works

DriftGuard has two main pipelines: recording and retrieval.

Recording pipeline

When you call guard.record(action, feedback, outcome):

action text
  ↓
normalize (spaCy lemmatize + stopword removal)
  ↓
embed (sentence-transformers)
  ↓
find similar existing node (cosine similarity > threshold)
  ↓
merge into existing node  OR  create new node
  ↓
add edges: action → feedback → outcome
  ↓
persist to JSON or SQLite

Semantic merge

If "add more salt" arrives after "increase salt" is already in the graph, the merge engine computes cosine similarity between the two normalized embeddings. If similarity exceeds the action threshold (default 0.72), the new text merges into the existing node — incrementing its frequency — rather than creating a duplicate.

This keeps the graph compact and ensures repeated paraphrased mistakes reinforce one signal rather than fragmenting it.

Retrieval pipeline

When you call guard.before_step(context):

context text
  ↓
embed
  ↓
top-k cosine similarity search over action nodes
  ↓
walk causal chains from each matched node (DFS)
  ↓
build warnings from (trigger, risk) pairs
  ↓
score confidence = f(node_freq, edge_freq, similarity, recency)
  ↓
deduplicate + sort by confidence descending
  ↓
return RetrievalResponse

The same pipeline runs a second time against the success graph, building reinforcements from (trigger, recommendation) pairs using the same confidence scoring. Both results — warnings and reinforcements — are attached to the same RetrievalResponse. See Success Memory.

Confidence scoring

confidence = 0.65 × reinforcement_score
           + 0.35 × similarity_score
           + recency_weight × recency_score

Reinforcement score tiers:

Combined frequency	Score
≥ 5	0.95
≥ 3	0.85
≥ 2	0.75
1	0.60

Graph structure

The in-memory graph is a directed networkx.DiGraph. Each node stores:

Attribute	Type	Purpose
type	str	action, feedback, outcome
embedding	np.ndarray	Normalized sentence embedding
frequency	int	How many times merged into
first_seen	datetime	When first recorded
last_seen	datetime	When last reinforced

Each edge stores frequency, weight, and created_at.

Pruning

Deep pruning runs on a schedule and removes:

1. Weak edges — frequency below prune_edge_min_frequency
2. Stale nodes — not seen within prune_node_stale_days
3. Isolated nodes — no incoming or outgoing edges after edge removal