Beyond Keywords and Vectors: DSL-Based Hybrid Search for Enterprise Code Discovery

Why Traditional Search Fails for Code

Engineering teams sit on millions of lines of code, RFCs, and runbooks. Keyword-only systems miss synonyms and architectural concepts; semantic-only systems ignore literal identifiers like `ORA-12154` or `getInvoiceStatus`. The downstream effect is endless context switching, duplicate implementations, and tribal knowledge that never scales.

• Pure semantic search understands concepts but overlooks exact error codes, function names, or API endpoints.
• Pure keyword search finds literals yet fails whenever authors write “exception handling” while you search for “error handling”.
• Most hybrid systems assume fixed weights, so the blend never adapts to the developer’s real intent.

DSL-Powered Hybrid Intelligence

CoderSwap generates a domain-specific policy at deployment time. The DSL encodes regex-based intent detectors alongside weight budgets for semantic, lexical, and recency signals. At query time the policy fires the first matching intent and streams safe bind parameters into Oracle—no dynamic SQL required.

{
  "intents": [
    {
      "pattern": "error|exception|ora-\\d+",
      "weights": { "semantic": 0.3, "lexical": 0.6, "recency": 0.1 }
    },
    {
      "pattern": "implement|create|build|develop",
      "weights": { "semantic": 0.7, "lexical": 0.2, "recency": 0.1 }
    }
  ],
  "defaults": { "semantic": 0.5, "lexical": 0.4, "recency": 0.1 }
}

Query-to-Result Architecture

Every search runs the same deterministic pipeline: detect intent, pull weights, execute vector and keyword branches in parallel, then fuse the scores with an intent-aware budget. A diversity filter keeps results from repeating the same file.

User Query
  ↓
Intent Detection (Regex)
  ↓
DSL Weight Selection
  ↓
Parallel Retrieval
    ├─ Vector Search (HNSW)
    └─ Keyword Search (Oracle TEXT)
  ↓
Weighted Fusion & Diversity Filter
  ↓
Results

WITH q AS (
  SELECT VECTOR_EMBEDDING(all_minilm_l12_v2 USING :query AS data) v FROM dual
)
SELECT
  chunk_id,
  text,
  metadata,
  (1 - VECTOR_DISTANCE(embedding, (SELECT v FROM q), COSINE)) AS semantic_sim,
  SCORE(1) AS lexical_sim,
  GREATEST(0, 1 - (MONTHS_BETWEEN(SYSDATE, created_at) / 36)) AS recency_score,
  (:w_semantic * semantic_sim +
   :w_lexical * lexical_sim +
   :w_recency * recency_score) AS final_score
FROM CS_PROJECT_CHUNKS
WHERE CONTAINS(text, :lexical_query, 1) > 0
   OR VECTOR_DISTANCE(embedding, (SELECT v FROM q), COSINE) < 0.8
ORDER BY final_score DESC
FETCH FIRST :k ROWS ONLY;

DSL Compilation Pipeline

The configuration lives entirely inside Oracle and is versioned for rollback. Anthropic Claude analyzes a 20K-token corpus sample, proposes a DSL policy, and CoderSwap validates weights, regex specificity, and schema before committing it.

def compile_weights(selected_weights):
    ALLOWED_IDENTIFIERS = {"semantic", "lexical", "recency"}

    terms, binds = [], {}
    for signal, weight in selected_weights.items():
        if signal in ALLOWED_IDENTIFIERS:
            param = f"w_{signal}"
            terms.append(f":{param} * {signal}")
            binds[param] = min(max(weight, 0.0), 1.0)

    return " + ".join(terms), binds

Future Directions

Why It Matters

Get Started in Minutes

Upload your repo, let CoderSwap build a DSL profile, and query your codebase with enterprise-grade safeguards.

# Upload your codebase
curl -X POST https://api.coderswap.ai/v1/upload \
  -H "X-API-Key: your-key" \
  -F "files=@src.zip"

# Generate search profile
curl -X POST https://api.coderswap.ai/v1/settings:generate \
  -H "X-API-Key: your-key" \
  -d '{"project_id": "your-project"}'

# Search with intelligence
curl -X POST https://api.coderswap.ai/v1/search \
  -H "X-API-Key: your-key" \
  -d '{"query": "ORA-12154 connection handling", "k": 10}'

References & Further Reading

• Hierarchical Navigable Small World (HNSW) Algorithm
• Oracle ADW Vector Search Documentation
• Anthropic Claude API
• BM25 Algorithm Explained
• Hybrid Search Best Practices
• SPLADE: Sparse Lexical and Expansion Model
• Cross-Encoder vs Bi-Encoder