Deep Dive
Reconciliation Logic
The Linkage Strategy (Pillar V): Balancing computational efficiency with semantic accuracy.
The Linkage Strategy (Pillar V) utilizes a two-step process to balance computational efficiency with semantic accuracy: separating initial candidate retrieval from final relevance adjudication.
3.1 Deep Dive: Retrieval vs. Adjudication
Requirement:
"Vendor must demonstrate experience with Oracle-to-Postgres database refactoring."
Phase A: Vector Triangulation
- Mechanism: Cosine Similarity Search
- Input: "Oracle", "Postgres", "Database", "Refactoring"
- Result: Retrieves ~20 candidates based on keyword similarity
Phase B: The Critic Agent
- Mechanism: LLM Reasoning
- Input: Candidates + Verbatim Requirement
- Logic: Filter candidates that do not explicitly mention migration from Oracle to Postgres
Adjudication Logic Table
| Candidate Snippet | Vector Match? | Critic Reasoning | Final Verdict |
|---|---|---|---|
| "Managed Oracle DBs for Army..." | YES (90%) | "Mentions Oracle, but no migration to Postgres." | DISCARD |
| "Migrated SQL Server to Postgres..." | YES (85%) | "Wrong source DB (SQL Server, not Oracle)." | DISCARD |
| "Refactored legacy Oracle 11g to Amazon RDS (Postgres)..." | YES (92%) | "Explicit match: Source=Oracle, Dest=Postgres." | KEEP |
The result of this two-pass process is a filtered set of evidence that meets the specific semantic logic of the requirement, rather than just matching keywords.