RAG Document Ingestion Pipeline

RAG Document Ingestion Pipeline

Build production-grade document ingestion pipelines that chunk, embed, and store documents in vector databases for retrieval-augmented generation.

Activation Triggers

Activate on: "document ingestion", "chunking strategy", "embedding pipeline", "vector DB ingestion", "RAG indexing", "ingest PDFs", "build knowledge base", "semantic chunking", "recursive chunking"

NOT for: LLM prompt design or retrieval query tuning (prompt-engineer, ai-engineer), vector DB operational migration (vector-database-migration-tool), or fine-tuning data preparation (fine-tuning-dataset-curator)

Quick Start

1. Identify sources — PDFs, HTML, Markdown, databases, APIs. Use unstructured or docling for parsing.
2. Choose chunking strategy — Recursive character splitting for general text, semantic chunking for domain-specific content, or document-structure-aware chunking for technical docs.
3. Select embedding model — text-embedding-3-large (OpenAI), embed-v4 (Cohere), or BAAI/bge-m3 (local). Match dimensionality to your vector DB plan.
4. Configure vector DB — Pinecone (managed), Qdrant (self-hosted or cloud), Weaviate (multi-tenant), or pgvector (Postgres-native).
5. Run ingestion with observability — Batch embed, upsert with metadata, validate retrieval quality on a test set.

Core Capabilities

Architecture Patterns

Pattern 1: Chunking Decision Tree

Document Type?
├── Structured (Markdown, HTML, code)
│   └── Structure-aware chunking (headers, functions)
│       └── Preserve hierarchy as metadata
├── Semi-structured (PDF with tables)
│   └── docling/unstructured → table extraction + text chunking
│       └── Embed tables as markdown, text as paragraphs
└── Unstructured (plain text, transcripts)
    └── Semantic chunking (embedding similarity breakpoints)
        └── Fallback: recursive character split (512-1024 tokens, 10% overlap)

Pattern 2: Production Ingestion Pipeline

Sources ──→ [Parser] ──→ [Chunker] ──→ [Enricher] ──→ [Embedder] ──→ [Vector DB]
  │            │            │              │              │              │
  │         unstructured  recursive/    add metadata:   batch embed   upsert with
  │         docling       semantic      source, date,   (batch=256)   namespace
  │                                     section, hash                 partitioning
  │
  └── Dedup by content hash before embedding (saves 30-50% cost)

# Production ingestion skeleton
from langchain_text_splitters import RecursiveCharacterTextSplitter
from hashlib import sha256

def ingest_documents(docs: list[str], collection: str):
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=512, chunk_overlap=64,
        separators=["\n\n", "\n", ". ", " "]
    )
    seen_hashes = set()
    chunks = []
    for doc in docs:
        for chunk in splitter.split_text(doc):
            h = sha256(chunk.encode()).hexdigest()[:16]
            if h not in seen_hashes:
                seen_hashes.add(h)
                chunks.append({"text": chunk, "hash": h})
    # Batch embed and upsert
    embeddings = embed_batch([c["text"] for c in chunks], batch_size=256)
    vector_db.upsert(collection, chunks, embeddings)

Pattern 3: Metadata-Enriched Chunks

Always store metadata alongside vectors for filtered retrieval:

metadata = {
    "source": "docs/api-reference.md",
    "section": "Authentication",
    "chunk_index": 3,
    "total_chunks": 12,
    "ingested_at": "2026-03-20T00:00:00Z",
    "content_hash": "a1b2c3d4",
    "token_count": 487,
}

Anti-Patterns

1. Fixed-size chunking without overlap — Splits mid-sentence, destroys context. Always use overlap (10-15%) or semantic boundaries.
2. Embedding everything without deduplication — Near-duplicate chunks waste storage and degrade retrieval. Hash-dedup or use MinHash for fuzzy dedup.
3. Ignoring chunk metadata — Without source, section, and date metadata, you cannot filter, cite, or refresh stale content.
4. One embedding model for all domains — Legal text and code have different semantic spaces. Benchmark retrieval quality per domain before committing.
5. No ingestion idempotency — Re-running ingestion creates duplicates. Use content hashes as vector IDs for upsert-based idempotency.

Quality Checklist

• Chunking strategy matches document structure (not just fixed-size)
• Chunk sizes benchmarked for retrieval quality (typically 256-1024 tokens)
• Overlap configured to preserve cross-boundary context
• Deduplication prevents redundant embeddings
• Metadata attached to every chunk (source, section, date, hash)
• Embedding model dimensionality matches vector DB index config
• Batch embedding with retry logic for API failures
• Ingestion is idempotent (re-run safe via content-hash IDs)
• Retrieval quality validated on test queries before production
• Monitoring: ingestion throughput, embedding latency, DB storage growth