Conversational memorial companion with zero-fabrication retrieval architecture
Technical Descriptor: "Zero-fabrication retrieval architecture" · "Structurally grounded, non-generative-fallback design"
Unlike generic RAG prototypes that patch hallucinations with prompt directives, Smriti closes fabrication structurally at the storage and retrieval boundary.
Interactive Execution Topology
Verified branch fork at Node 1, asymmetric tone latching at Node 4, and 600ms resilience backoff at Node 6.
Bounded Memory & Cosine Similarity
Computes top-5 cosine similarity over memory embeddings directly in-process (zero network round-trip). Bounded vault design eliminates distributed vector database latency overhead.
Rule-Based Tone Classifier
Regex-driven tone categorization evaluates synchronously in-process (zero LLM overhead), bypassing model classification nondeterminism and guaranteeing grief-appropriate conversational warm grounding.
Zero Unconfirmed Firestore Writes
All memory updates require explicit client UI confirmation (Save / Review / Dismiss). Prevents cross-turn hallucination accumulation from entering the permanent database.
Deterministic 503 Exponential Backoff
600ms backoff interval on upstream model unavailability, automatically rerouting to fallback models without dropping conversational state or breaking SSE stream buffers.