agentdb-vector-search

L1 Improvement

• Refactored into the Skill Forge required sections with clear triggers, contracts, and validation for AgentDB search flows.
• Added prompt-architect ceilings, constraint capture, and integration checklists for vector search deployments.

STANDARD OPERATING PROCEDURE

Purpose

Guide teams to implement semantic search using AgentDB vector search with correct indexing, query tuning, and evaluation practices.

Trigger Conditions

• Positive: building or optimizing semantic search, troubleshooting retrieval quality, or selecting embeddings/index params.
• Negative/reroute: non-AgentDB search stacks, unrelated prompt tuning (prompt-architect), or database persistence patterns (agentdb-memory).

Guardrails

• Ensure embeddings and metadata schemas are defined before ingest.
• Enforce evaluation with relevance metrics (nDCG/MRR) and adversarial queries.
• Keep outputs in English with explicit confidence ceilings; document parameter choices.
• Avoid overfitting to small eval sets; prefer reusable harnesses.

Execution Phases

1. Scoping: Capture corpus type, domains, latency targets, and constraints; classify HARD/SOFT/INFERRED.
2. Index Design: Choose embedding model, dimension, metadata, and filters; set up ingestion plan.
3. Query Strategy: Define similarity metrics, top-k, filters, and rerankers; plan for miss cases.
4. Evaluation: Build/query eval sets (happy/edge/adversarial); measure precision/recall-like metrics.
5. Delivery: Document configuration, sample queries, failure handling, and maintenance routines.

Pattern Recognition

• Short text search → prioritize fast embeddings and higher k with reranking.
• Long-form docs → chunking strategy with overlap and metadata facets.
• Safety-critical domains → add blocklists and human-in-the-loop review for low-confidence hits.

Advanced Techniques

• Hybrid retrieval combining dense + keyword filters.
• Dynamic top-k based on query entropy or user tier.
• Logging-based feedback loops feeding recursive-improvement and reindexing.

Common Anti-Patterns

• Ingesting without schema leading to noisy metadata.
• No evaluation harness; relying on spot checks.
• Ignoring miss cases or silent failures.

Practical Guidelines

• Version embedding models and indexes; record config hashes.
• Include retry/backoff for upstream embedding services.
• Document cold-start guidance and reindex cadence.

Cross-Skill Coordination

• Upstream: prompt-architect for query clarity; base-template-generator for scaffolds.
• Parallel: agentdb-optimization for tuning; agentdb-advanced for complex filters; recursive-improvement for ongoing evaluation.
• Downstream: agent-creator embedding search behaviors into agents.

MCP Requirements

• Requires AgentDB vector search MCP; tag WHO=agentdb-vector-search-{session}, WHY=skill-execution for memory traces.

Input/Output Contracts

inputs:
  corpus_description: string  # required
  embedding_model: string  # optional preference
  constraints: list[string]  # optional performance/safety constraints
outputs:
  search_plan: file  # index/query configuration and rationale
  eval_plan: file  # test cases, metrics, and results
  runbook: summary  # monitoring and maintenance steps

Recursive Improvement

• Feed retrieval errors and user feedback into recursive-improvement to retune chunking, filters, or scoring weights.

Examples

• Configure semantic search for product docs with metadata facets and hybrid retrieval.
• Tune Q&A search for support tickets with adversarial queries covering ambiguous intents.

Troubleshooting

• Low relevance → adjust chunking, filters, and rerankers; expand eval set.
• Latency spikes → tune top-k and prefiltering; cache frequent queries.
• Skewed results → review embedding choice and retrain on in-domain data if available.

Completion Verification

• Index/query design documented with parameters and rationale.
• Evaluation harness executed; metrics and ceilings recorded.
• Safety/edge handling noted (miss cases, blocklists, fallbacks).
• Monitoring and maintenance steps delivered.

Confidence: 0.70 (ceiling: inference 0.70) - AgentDB vector search SOP rewritten with Skill Forge cadence and prompt-architect ceilings.