Pipeline Flow

Pipeline Flow

1. Input Guardrail

Purpose:

• Prompt injection detection
• Unsafe content detection
• Malicious query filtering

Methods:

• Pattern Matching
• LLM Safety Validation

---

2. Query Classification

Determines whether a query belongs to the knowledge domain.

Output:

question_is_relevant: bool

---

3. Hybrid Retrieval

Combines:

• Semantic Search (ChromaDB)
• Keyword Search (BM25)

---

4. Reciprocal Rank Fusion (RRF)

Merges rankings from:

• ChromaDB
• BM25

Produces:

retrieved_docs

---

5. Context Guardrail

Checks context size.

Maximum:

max_context_tokens = 1000

Uses:

tiktoken

---

6. Generate Node

Model:

gpt-4o-mini

Retries:

3 attempts

Backoff:

Exponential

---

7. Output Guardrail

• PII Redaction
• Safety Validation

---

8. Confidence Scoring

Self-RAG style confidence estimation.

Range:

0.0 - 1.0