BCI in Amprealize

BCI (Behavior-Conditioned Inference) in Amprealize

What It Is

BCI is Amprealize's implementation of RAG for procedural knowledge. Instead of retrieving documents, it retrieves behaviors — proven step-by-step strategies that condition the agent's execution.

How It Maps to Concepts

AI/ML Concept	Amprealize Implementation
Embeddings	`behavior_retriever.py` embeds behavior descriptions via sentence-transformers
FAISS	In-process FAISS index stores behavior vectors for fast similarity search
RAG	Query → retrieve top-k behaviors → inject into prompt → generate
Hybrid Retrieval	`advanced_retrieval_service.py` combines FAISS vector search with keyword filtering

The Flow

Agent receives task
    ↓
behaviors.getForTask(task_description, role)
    ↓
[1] Embed task description → query vector
    ↓
[2] FAISS search → top-k behavior candidates
    ↓
[3] Filter by role (Student/Teacher/Strategist)
    ↓
[4] Score and rank by relevance
    ↓
[5] Return behavior set → injected into agent context
    ↓
Agent executes with behavior conditioning

Key Files

amprealize/bci_service.py — Core BCI orchestration
amprealize/behavior_retriever.py — Embedding + FAISS retrieval
amprealize/advanced_retrieval_service.py — Hybrid retrieval with re-ranking
amprealize/context_composer.py — Assembles behaviors into prompts

Why BCI Over Fine-Tuning

Approach	Pros	Cons
BCI (RAG)	Instant updates, no training cost, auditable	Depends on retrieval quality
Fine-tuning	Baked into model weights	Expensive, slow to update, opaque

Amprealize chose BCI because behaviors change frequently (new patterns discovered weekly) and auditability matters (every behavior citation is traceable).