Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/langchain-py-pack/skills/langchain-data-handling/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill langchain-data-handling. Copyright stays with the author (MIT).

LangChain Data Handling — Loaders and Splitters (Python)

Overview

You have a RAG system over a Python docs site. A user asks "what does trim_messages do?" and the retriever returns this chunk:

### `trim_messages(strategy="last", include_system=True)`

Trim a message history to fit a token budget. The newest messages are kept;
older messages are dropped. Pass `include_system=True` to preserve the system

...and that's it. The chunk ends there. The code example showing the function body — the actual thing the user wanted — is in a different chunk, retrieved with a lower similarity score and dropped before the LLM sees it. The model then hallucinates the function's behavior from the signature alone.

This is pain-catalog entry P13. RecursiveCharacterTextSplitter's default separators are ["\n\n", "\n", " ", ""]. It splits on any blank line — including inside triple-backtick code fences in Markdown. The fix is a one-line swap to RecursiveCharacterTextSplitter.from_language(Language.MARKDOWN), which treats the fence as an atomic unit, but you have to know the bug exists.

The sibling failures this skill prevents:

• P49 — PyPDFLoader splits by page. A 5-row financial table that spans a page break gets torn in half; rows 1-3 go in one chunk, rows 4-5 in another with no header. A RAG answer sourced from the second chunk misquotes the numbers because the column meanings are in the first chunk. Fix: use PyMuPDFLoader or UnstructuredPDFLoader, which detect tables and emit them as distinct structured elements.
• P50 — WebBaseLoader's default User-Agent is python-requests/2.x. Cloudflare-protected sites flag this as a bot and return a 403 interstitial HTML page ("Checking your browser...") instead of real content. The crawler indexes the challenge page. You notice weeks later when every retrieval from that source returns the same Cloudflare text. Fix: set a realistic header_template={"User-Agent": "Mozilla/5.0 ..."}, respect robots.txt, and rate-limit per-host to 1 req/sec.

Pinned versions: langchain-core 1.0.x, langchain-community 1.0.x, langchain-text-splitters 1.0.x, pymupdf, unstructured. Pain-catalog anchors: P13, P49, P50, P15.

This skill is the upstream half of the RAG pipeline — load and chunk. For the downstream half (embedding, scoring, reranking) see the pair skill langchain-embeddings-search, which covers score semantics (P12), dim guards (P14), and reranker filtering (P15). Do not re-implement chunking there.

Prerequisites

• Python 3.10+
• langchain-core >= 1.0, < 2.0 and langchain-community >= 1.0, < 2.0
• langchain-text-splitters >= 1.0, < 2.0
• PDF support: pip install pymupdf unstructured[pdf]
• Web loading: pip install beautifulsoup4 requests
• For corpus dedup (optional): pip install datasketch

Instructions

Step 1 — Choose a loader by source format

Loader selection is the first decision — get it wrong and no amount of splitter tuning will recover. Use the decision table:

from langchain_community.document_loaders import (
    PyMuPDFLoader,            # table-aware PDF
    WebBaseLoader,            # web pages (set custom UA)
    UnstructuredMarkdownLoader,
    DirectoryLoader,
)

# PDF with tables — P49 fix
pdf_docs = PyMuPDFLoader("10-Q-filing.pdf").load()

# Web page — P50 fix
web_docs = WebBaseLoader(
    "https://example.com/article",
    header_template={
        "User-Agent": "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36"
    },
).load()

# Markdown docs site
md_docs = UnstructuredMarkdownLoader("docs/guide.md").load()

# Corpus
corpus = DirectoryLoader(
    "./docs", glob="**/*.md",
    loader_cls=UnstructuredMarkdownLoader,
    show_progress=True,
).load()

Hard limit: keep single-PDF ingestion under 5 MB per call. Larger files should be pre-split with pdftk / qpdf to avoid OOM on PyMuPDFLoader's full-document parse.

See Loader Selection Matrix for the full per-format table with cost and accuracy notes.

Step 2 — Pick a splitter by content type

from langchain_text_splitters import (
    RecursiveCharacterTextSplitter,
    Language,
    HTMLHeaderTextSplitter,
)

# GOOD — P13 fix for Markdown
md_splitter = RecursiveCharacterTextSplitter.from_language(
    Language.MARKDOWN, chunk_size=1000, chunk_overlap=100,
)

# GOOD — Python code
py_splitter = RecursiveCharacterTextSplitter.from_language(
    Language.PYTHON, chunk_size=1500, chunk_overlap=150,
)

# GOOD — HTML long-form with heading-as-metadata
html_splitter = HTMLHeaderTextSplitter(
    headers_to_split_on=[("h1", "Header 1"), ("h2", "Header 2")],
)

# BAD — breaks inside code fences (P13)
bad = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)

See Language-Aware Splitters for the full list of Language.* enum values, custom separator patterns, and the code-fence-detection regex for when you need a custom splitter.

Step 3 — Tune chunk_size and overlap

Defaults from the table work for most corpora. Tune when:

• Retrieval misses context: increase chunk_size (1000 → 1500) or chunk_overlap (100 → 200). Overlap is what bridges a concept that crosses chunk boundaries.
• Retrieval too broad, answers wander: decrease chunk_size (1000 → 500). Smaller chunks = more precise retrieval but more chunks to index.
• Tables / structured data: do NOT tune — index them separately (step 4).

A 1% overlap-to-size ratio is too low (200/20000); 20% is the sweet spot for most prose. Code needs less overlap (10%) because function boundaries are natural splits.

Step 4 — Detect and index tables as structured records

Tables are not text. If your corpus has financial filings, product specs, or any tabular data, index tables as separate records with column metadata:

import fitz  # pymupdf directly for table detection

def extract_tables_as_records(pdf_path: str) -> list[dict]:
    """Extract tables as one record per row."""
    doc = fitz.open(pdf_path)
    records = []
    for page_num, page in enumerate(doc):
        tables = page.find_tables()
        for table in tables:
            rows = table.extract()
            if not rows:
                continue
            headers = rows[0]
            for row_idx, row in enumerate(rows[1:], start=1):
                record = {
                    "page": page_num,
                    "table_idx": tables.tables.index(table),
                    "row_idx": row_idx,
                    "content": " | ".join(f"{h}: {v}" for h, v in zip(headers, row)),
                    "metadata": dict(zip(headers, row)),
                }
                records.append(record)
    return records

Now a question like "what was Q3 revenue?" retrieves a single row with its column headers attached, not half a table missing the column meanings. See Table Preservation for the full pattern including hybrid retrieval (prose + table records).

Step 5 — Preserve metadata through the pipeline

The loader attaches metadata (source, page, heading); the splitter propagates it. Front-matter in Markdown, PDF page numbers, and web URLs should all end up in doc.metadata so retrieval results are citable:

for doc in md_docs:
    # Markdown front-matter (if loader extracted it)
    print(doc.metadata.get("title"), doc.metadata.get("date"))

# Splitter-preserved metadata
chunks = md_splitter.split_documents(md_docs)
assert chunks[0].metadata == md_docs[0].metadata  # preserved

Custom metadata (tenant_id, version, confidence) should be added before splitting so every chunk inherits it.

Step 6 — Deduplicate noisy corpora

Web crawls and scraped docs often contain near-duplicate pages (nav chrome, footer boilerplate, syndicated posts). MinHash-based dedup at the chunk level keeps the index clean:

from datasketch import MinHash, MinHashLSH

lsh = MinHashLSH(threshold=0.9, num_perm=128)
kept = []
for i, chunk in enumerate(all_chunks):
    mh = MinHash(num_perm=128)
    for tok in chunk.page_content.lower().split():
        mh.update(tok.encode())
    if not list(lsh.query(mh)):
        lsh.insert(str(i), mh)
        kept.append(chunk)

A threshold of 0.9 catches near-duplicates (minor wording differences) without eating legitimate paraphrases.

Step 7 — Compose the pipeline

# Multi-stage: load → split → dedup → index
def build_rag_index(source_dir: str, store):
    # 1. Load
    docs = DirectoryLoader(
        source_dir, glob="**/*.md",
        loader_cls=UnstructuredMarkdownLoader,
    ).load()

    # 2. Clean (empty-content filter)
    docs = [d for d in docs if d.page_content.strip()]

    # 3. Split (language-aware)
    splitter = RecursiveCharacterTextSplitter.from_language(
        Language.MARKDOWN, chunk_size=1000, chunk_overlap=100,
    )
    chunks = splitter.split_documents(docs)

    # 4. Dedup (optional for noisy corpora)
    # chunks = dedup_minhash(chunks, threshold=0.9)

    # 5. Index — handoff to langchain-embeddings-search
    store.add_documents(chunks)
    return store

For the embedding + indexing + retrieval steps, see langchain-embeddings-search.

Output

• Loader chosen from the selection matrix matching source format and table needs
• Splitter chosen from the decision tree matching content type
• Chunk size + overlap tuned from the defaults (1000/100 prose, 1500/150 code, 500/50 FAQ)
• Tables extracted as structured records with column metadata (not text chunks)
• Web loaders configured with realistic User-Agent and robots.txt respect
• Metadata preserved through loader → splitter → index
• Optional MinHash dedup (threshold 0.9) for noisy corpora

Error Handling

Examples

Ingesting a Markdown docs site with code examples

Markdown docs with Python code fences require Language.MARKDOWN to keep fence boundaries intact. Chunk size 1000 with 100 overlap preserves one function-sized example per chunk. Front-matter fields (title, date, author) are attached as metadata for citation. See Language-Aware Splitters.

Ingesting a PDF filing with financial tables

10-Q filings have dozens of multi-row tables. Use PyMuPDFLoader for the prose and a direct fitz.find_tables() pass to extract tables as structured records. Index prose with chunk_size=1000 and tables as one-row-per-record with the header row concatenated. Questions like "what was Q3 revenue?" hit a single row with column meanings attached. See Table Preservation.

Crawling a documentation site behind Cloudflare

Set a realistic User-Agent, fetch robots.txt first and respect Disallow rules, rate-limit to 1 req/sec per host, and prefer the site's sitemap or RSS feed when available. Assert response length > 500 chars and reject known interstitial patterns. See Crawler Hygiene.

Ingesting a Python code repo for code RAG

GenericLoader with LanguageParser(language=Language.PYTHON) preserves function and class boundaries. Chunk size 1500 with 150 overlap gives enough context for typical function-level queries. Imports and module docstrings end up in their own chunks — tag them with metadata for higher precision retrieval on "where is X imported from" queries.

Resources

• LangChain Python: Text splitters
• LangChain Python: Document loaders
• LangChain Python: Retrievers
• PyMuPDF (fitz) docs — table detection via page.find_tables()
• Unstructured.io — table-aware PDF parsing
• Cloudflare bot management — why default UAs fail
• robots.txt spec — crawler obligations
• Pair skill: langchain-embeddings-search (embed/score/rerank — downstream of this skill)
• Pack pain catalog: docs/pain-catalog.md (entries P13, P49, P50, P15)