Page 398 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

• Convert the input name to a dense embedding
• Precompute embeddings for existing names in the database and store them in:

• Elasticsearch vector index (with dense_vector)

5) Semantic Search
• Use cosine similarity between input name and stored name embeddings
• Retrieve top N semantically similar names

6) Filter Results

• Remove any names that caused rejection (e.g., exact match or prohibited)
• Filter based on policy: sensitive words, reserved names, etc.
• Prioritize names that:

• Are semantically similar but legally and contextually acceptable
• Match user’s language preference

7) Scoring & Ranking
• Combine scores:

• Rule-based score (linguistic similarity, uniqueness, policy compliance)
• SBERT score (semantic similarity)
• Weighted average or customizable scoring strategy

8) Response Generation: Return a structured response

362