TECHNICAL SPECIFICATION V1.0

ARCHITECTURE & SPECS //

System architecture, vector indexing schemas, and machine-readable agent protocol surfaces.

1. Multimodal Vision Pipeline

When an image is ingested, Pomniter runs a two-stage parsing pass:

Layer A (OCR Extract): High-precision text localization for exact tokens (code keywords, flight numbers, dates).
Layer B (Multimodal Reasoning): Vision model analyzes layout, syntax indentation, diagram flow, and conceptual domain to generate a rich natural language summary.

2. Hybrid Vector Indexing

Pomniter utilizes a hybrid retrieval mechanism:

Score = (α · CosineSimilarity(q_vec, d_vec)) + ((1 - α) · BM25(q_tokens, d_text))

This guarantees that fuzzy concept queries match via vector semantics while exact keywords (like ticket numbers or variable names) are preserved via BM25 lexical search.

3. Metadata JSON Schema

{
  "id": "mem_8941fba0",
  "created_at": 1756080000,
  "category": "programming",
  "topic": "binary_search",
  "summary": "Midpoint overflow bug fix",
  "entities": ["mid", "left", "right"],
  "tags": ["#LeetCode", "#DSA"],
  "volatility": "persistent",
  "vector_dim": 1536
}

IS-AGENTIC.COM COMPLIANCE

Agent Discovery Surfaces

Pomniter exposes standardized machine-readable endpoints so that AI agents and search engines can discover, query, and cite content with zero friction.

Inspect /llms.txt Inspect /llms-full.txt Inspect /.well-known/agent.json

Local-First Privacy

All vector embeddings can be generated and queried locally on device (Android/Desktop) with AES-256 encrypted database vaults. No personal screenshots are ever used for public model training.