The bridge between
business and code.
“Customer Onboarding” and process_kyc_check() share zero common language.
Standard search fails completely.
Hypothetical Document Embedding bridges the gap.
Mathematically.
Why Standard Search Fails
The vocabulary gap is structural.
Business people speak in concepts: “Know Your Customer,” “Loan Origination,” “Risk Assessment.” Engineers write functions: validate_identity(),calc_debt_service_ratio().
Keyword search returns nothing. Embedding similarity returns noise.
The domains are too far apart in vector space.
Matching attempts
The Mechanism
Two directions. One truth.
Forward HyDE
1. Take a business concept
2. LLM generates 5 hypothetical code snippets
3. Compare synthetic code to real code via cosine similarity
→ Code matches code. Domain gap eliminated.
Reverse HyDE
1. Take real code chunks
2. LLM generates hypothetical business descriptions
3. Compare synthetic descriptions to real concept definitions
→ Business language matches business language.
Both directions run.
Results are combined with weighted scoring.
Every link gets a confidence score.
Ambiguous matches are flagged for human review, not silently accepted.
The economics.
$0.50
per 15 concepts (one-time)
$5
/month at 50K chunks
+25%
recall improvement
This site uses cookies
We use essential cookies for the site to function and analytics cookies (Google Analytics) to understand how you use it. Analytics cookies are only activated with your consent. We do not track you across other websites. Your data is stored in the EU and processed in accordance with GDPR. Read our Privacy Policy
