如何用 Pinecone + Claude（Python、SDK v10）為公司文件建立問答機械人

AI disclosure: This article was written with the help of AI and reviewed by me. Every API call in it was checked against the SDK source of pinecone==10.3.0 and anthropic==2.4.0, and prices and limits were taken from the official pages linked inline (as of November 2026). Treat the code as a starting point: run it against your own keys before relying on it.

What this solves

Once a company has enough internal docs, the same problems show up: new hires can't find the leave policy, nobody remembers which version of the expense rules is current, and wiki search only works if you guess the exact words the author used ("PTO" won't find the page that says "annual leave").

Retrieval-augmented generation (RAG) is the usual fix: split documents into chunks, embed them into a vector database, retrieve the most relevant chunks for each question, and have an LLM answer only from those chunks, with citations.

This walkthrough builds a command-line Q&A bot with Pinecone (managed vector database with hosted embedding and rerank models) and Claude (Anthropic's LLM, for answer generation):

1. Chunking
2. Embedding with Pinecone's integrated embedding (text is vectorized on write)
3. Creating a serverless index and upserting
4. Retrieval + reranking
5. Prompting Claude for a cited answer
6. The full script, dependencies and env vars
7. Cost estimates and common pitfalls

Versions used (latest on PyPI in November 2026):

The Pinecone SDK changed a lot in v10. Many tutorials still use the v5–v7 style (create_index_for_model, upsert_records, ServerlessSpec), which may fail or emit deprecation warnings. The pitfalls section maps old calls to new ones.

1. Architecture

         ┌──────────── Offline ingest ─────────────────┐
 .md/.txt ──> chunk ──> documents.upsert ──> Pinecone index
                              │              (llama-text-embed-v2
                              │               embeds on write)
         └─────────────────────────────────────────────┘

         ┌──────────── Online ask ─────────────────────┐
 question ──> documents.search (embed scoring, recall 12)
          ──> inference.rerank (bge-reranker-v2-m3, keep 4)
          ──> build prompt ──> Claude Messages API ──> answer with [n] citations
         └─────────────────────────────────────────────┘

Embedding, storage and reranking all happen on Pinecone's side, so you don't install PyTorch or host an embedding model. The trade-off: your text is processed by Pinecone's hosted models (more on that under data residency).

2. Setup

2.1 Pinecone API key

Create an account at pinecone.io, then create a key on the API Keys page of the console (https://app.pinecone.io). The free Starter plan is enough for everything here, but note its limits (as of November 2026, see Pricing and Rate limits):

• Indexes can only be created in AWS us-east-1.
• Up to 5 serverless indexes per project, 100 namespaces per index, 2 GB storage per organization.
• 1M read units, 2M write units and 1 GB egress per month.
• 5M tokens per month per hosted embedding model.
• bge-reranker-v2-m3: 500 requests per month, 60 per minute. This runs out fast, so the script has a switch to turn reranking off.

2.2 Anthropic API key

Create a key in the Claude Console (https://platform.claude.com). Check Anthropic's supported countries list for your region first.

2.3 Dependencies

requirements.txt:

pinecone==10.3.0
anthropic==2.4.0

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

The package is pinecone, not the old pinecone-client.

2.4 Environment variables

export PINECONE_API_KEY="your-pinecone-key"
export ANTHROPIC_API_KEY="your-anthropic-key"

Don't hard-code keys or commit them to Git.

3. Step 1: Chunking

Chunking is the first thing that makes or breaks RAG quality. Chunks that are too large mix several topics, so the embedding gets "averaged out" and you waste prompt tokens. Chunks that are too small lose context and the model gets half a sentence.

A simple strategy that works for Markdown or plain-text policy docs:

1. Split on blank lines into paragraphs, avoiding mid-paragraph breaks.
2. Pack adjacent paragraphs into chunks of at most 500 characters.
3. Hard-split any single paragraph longer than 500 characters with a sliding window.
4. Keep an 80-character overlap between chunks so an answer is less likely to fall across a boundary.

def split_text(text: str, size: int = 500, overlap: int = 80) -> list[str]:
    paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
    chunks, buf = [], ""
    for p in paragraphs:
        if len(p) > size:
            if buf:
                chunks.append(buf)
                buf = ""
            step = size - overlap
            chunks.extend(p[i:i + size] for i in range(0, len(p), step))
            continue
        if len(buf) + len(p) + 1 <= size:
            buf = f"{buf}\n{p}" if buf else p
        else:
            chunks.append(buf)
            buf = (buf[-overlap:] + "\n" + p) if overlap else p
    if buf:
        chunks.append(buf)
    return chunks

This counts characters, not tokens, to avoid depending on a tokenizer. 500 characters of English is roughly 100–150 tokens, well under both limits that matter here:

• llama-text-embed-v2: max 2048 tokens per input, truncated by default beyond that (Embedding models).
• bge-reranker-v2-m3: query + document combined max 1024 tokens (Rerank results).

For finer strategies (by sentence, by token, by heading), see Pinecone's Chunking Strategies.

Each chunk carries metadata: source (relative path), category (top-level folder such as hr/ or finance/) and chunk_index. category is used for filtering, source for citations and for cleaning up on re-ingest.

4. Step 2: Choosing an embedding approach

Anthropic doesn't offer an embedding model (Embeddings), so you need one from somewhere else. Three common options:

Pinecone-hosted dense embedding models (docs as of November 2026):

On the Standard plan, llama-text-embed-v2 is listed at $0.16 per million tokens and multilingual-e5-large at $0.08 (Pricing).

This walkthrough uses llama-text-embed-v2: multilingual, a roomier input limit than e5, and available on Starter. If your docs are mostly short Q&A pairs, try multilingual-e5-large: change the model name in the schema and rebuild the index.

5. Step 3: Creating a serverless index

In SDK v10 the recommended index type is the document index. You declare the scoring fields with SchemaBuilder; every other field you write is stored as metadata and indexed for filtering automatically. Adding an embed config to a string field makes Pinecone embed that field's text on write, and embed the query text at search time (Integrated embedding).

import os
from pinecone import Pinecone, SchemaBuilder

pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"])

schema = (
    SchemaBuilder()
    .add_string_field(
        name="chunk_text",
        embed={"model": "llama-text-embed-v2", "dimension": 1024},
    )
    .build()
)

if not pc.indexes.exists(name="company-docs"):
    pc.indexes.create(
        name="company-docs",
        schema=schema,
        deployment={"deployment_type": "managed", "cloud": "aws", "region": "us-east-1"},
    )

Notes:

• deployment_type: "managed" means serverless. In v10, spec=ServerlessSpec(...) is deprecated (the SDK source says it's translated into deployment). If you omit deployment, the SDK defaults to a managed index in AWS us-east-1; it's spelled out here so you know where to change the region.
• pc.indexes.create blocks until the index is ready by default. Pass timeout=-1 to return immediately.
• Only declare scoring fields in the schema. Putting metadata like source or category in it returns a 400 at creation.
• The schema can't be changed after creation, including the embedding model and dimension. Switching models means deleting and rebuilding the index.
• An embed field also stores the original text and enables full-text search by default (store_text defaults to true), so search results can return chunk_text. Set store_text=False if you don't want the raw text stored in Pinecone, but then you'll have to look it up by ID in your own storage.

6. Step 4: Upserting documents

index = pc.index(name="company-docs")

documents = [
    {
        "_id": "3be9998a6b0d7b83#0",          # ASCII only
        "chunk_text": "Employees get 5 days of paid annual leave after one year of service...",
        "source": "hr/annual-leave.md",       # metadata, filterable automatically
        "category": "hr",
        "chunk_index": 0,
    },
    # ...
]

for i in range(0, len(documents), 96):
    index.documents.upsert(namespace="default-tenant", documents=documents[i:i + 96])

Key points:

• Indexes with an embed field accept at most 96 documents per upsert (1000 for regular indexes), because each request is embedded in a single call.
• Document _id must be 1–512 ASCII characters. Non-ASCII file names (accents, CJK) will fail SDK validation, so the script uses the first 16 hex chars of the path's SHA-1 plus the chunk number.
• Each document (including metadata) is capped at 2 MB; the embed field text at 100 KB; metadata at 40 KB.
• Indexing is asynchronous: a query right after a write may miss it. The official quickstart also sleeps a few seconds before querying.
• usage.embed_total_tokens in the response tells you how many embedding tokens the write used, which helps reconcile costs.

Re-ingesting: if a file shrinks and produces fewer chunks, the old extra chunks linger. The script calls documents.delete(filter=...) by source before writing, to clear a file's old chunks.

7. Step 5: Retrieval + rerank

7.1 Vector recall

resp = index.documents.search(
    namespace="default-tenant",
    top_k=12,
    score_by=[{"type": "embed", "fields": ["chunk_text"], "query": "How much annual leave do I get?"}],
    include_fields=["chunk_text", "source", "chunk_index"],
    filter={"category": {"$eq": "hr"}},     # optional metadata filter
)
for m in resp.matches:
    print(m.id, m.score, m.get("source"))

• With embed scoring, Pinecone embeds the query with the model fixed at index creation, so write and query always use the same model and dimension.
• Without include_fields you only get _id and _score. This is a classic beginner trap: the search "works" but returns no text. ["*"] returns every field.
• An embed clause must appear alone in score_by; it can't be combined with a BM25 text clause. For hybrid search, the docs suggest running both searches and merging with RRF (Integrated embedding).

7.2 Why rerank

Vector recall uses a bi-encoder: question and documents are embedded separately and compared by distance. It's fast but coarse. A rerank model (cross-encoder) scores the question and each document together, which is more accurate but slower. The usual pattern is to over-recall with vectors (12 here) and rerank down to the best few (4 here) for the prompt.

ranked = pc.inference.rerank(
    model="bge-reranker-v2-m3",
    query="How much annual leave do I get?",
    documents=[{"id": m.id, "chunk_text": m.get("chunk_text", "")} for m in resp.matches],
    rank_fields=["chunk_text"],
    top_n=4,
    return_documents=False,
    parameters={"truncate": "END"},
)
for r in ranked.data:
    print(r.index, r.score)   # r.index is the position in the documents you passed in

Model choice: Pinecone currently hosts three rerank models: bge-reranker-v2-m3 (multilingual, available on Starter), cohere-rerank-4-fast (Standard and above, processed via Azure, possibly outside the US), and pinecone-rerank-v0 (deprecated; don't use it in new projects).

bge-reranker-v2-m3 defaults to truncate: NONE, which errors on over-long input, so END is set explicitly.

Older tutorials often do recall + rerank in one call via index.search(..., rerank={...}). That's for legacy integrated indexes (created with create_index_for_model, using the Records API). With document indexes, rerank is a separate pc.inference.rerank call.

8. Step 6: Prompting Claude

Wrap each retrieved chunk in numbered XML tags so the model has anchors to cite:

SYSTEM_PROMPT = """You are an internal knowledge-base assistant. Answer employees' questions using ONLY the material inside <documents>.
Rules:
1. If the answer isn't in the material, say "I couldn't find this in the knowledge base." Don't make anything up.
2. After each key statement, cite the source as [n], where n matches <document index="n">.
3. If sources conflict, list the conflict instead of picking one.
4. Answer concisely."""

def build_user_message(question, hits):
    blocks = "\n".join(
        f'<document index="{i}" source="{h["source"]}">\n{h["text"]}\n</document>'
        for i, h in enumerate(hits, start=1)
    )
    return f"<documents>\n{blocks}\n</documents>\n\nQuestion: {question}"

Calling the Messages API:

import anthropic

client = anthropic.Anthropic()   # reads ANTHROPIC_API_KEY

response = client.beta.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=16000,
    system=SYSTEM_PROMPT,
    messages=[{"role": "user", "content": build_user_message(question, hits)}],
    output_config={"effort": "medium"},
    fallbacks="default",
    betas=["server-side-fallback-2026-07-01"],
)

if response.stop_reason == "refusal":
    print("The model declined to answer.")
else:
    print("".join(b.text for b in response.content if b.type == "text"))

What's different from older code:

• response.content is a list of blocks, not necessarily a single text block. Claude Sonnet 5.5 uses adaptive thinking by default, so the response can include thinking blocks. Filter on type == "text" instead of reading content[0].text.
• output_config.effort controls reasoning depth and token spend: low / medium / high / xhigh / max. Sonnet 5.5 defaults to high. Grounded Q&A rarely needs deep reasoning, so medium is used here; compare low and medium on your own question set.
• Don't pass temperature. API version 2026-10-21, the default in Anthropic SDK 2.x, rejects any request containing temperature, top_p or top_k with a 400. Delete the temperature=0 that old code often has.
• stop_reason == "refusal": Sonnet 5.5 has safety classifiers, and a refusal comes back as a normal response, not an exception. fallbacks="default" plus the beta header server-side-fallback-2026-07-01 lets the server retry a refused request on a recommended fallback model (Refusals and fallback). Because it's a beta, it goes through client.beta.messages.create. To skip the beta, drop the last two arguments and use client.messages.create; keep the refusal check either way.
• max_tokens is a cap, not a prepayment; you're billed for actual output. Set it too low and answers get cut off.

9. The full script

Save as rag_bot.py:

"""
rag_bot.py - company-docs Q&A bot on Pinecone (integrated embedding + rerank) + Claude

Usage:
    python rag_bot.py ingest ./docs                        # chunk and upsert
    python rag_bot.py ask "How much annual leave do I get?"
    python rag_bot.py ask "expense process" --category finance   # search one category only

Versions at time of writing: pinecone==10.3.0, anthropic==2.4.0, python>=3.10
"""

import argparse
import hashlib
import os
import sys
import time
from pathlib import Path

import anthropic
from pinecone import Pinecone, SchemaBuilder
from pinecone.errors import NotFoundError

# ---------- config ----------
INDEX_NAME = os.environ.get("PINECONE_INDEX", "company-docs")
NAMESPACE = os.environ.get("PINECONE_NAMESPACE", "default-tenant")
EMBED_MODEL = "llama-text-embed-v2"      # Pinecone-hosted embedding model
EMBED_DIM = 1024                          # this model supports 384/512/768/1024/2048
RERANK_MODEL = "bge-reranker-v2-m3"      # Pinecone-hosted rerank model (multilingual)
USE_RERANK = os.environ.get("USE_RERANK", "1") == "1"
CLAUDE_MODEL = "claude-sonnet-5-5"

CHUNK_SIZE = 500        # target characters per chunk
CHUNK_OVERLAP = 80      # overlap between adjacent chunks
UPSERT_BATCH = 96       # max docs per upsert for indexes with an embed field
RECALL_TOP_K = 12       # vector recall size
FINAL_TOP_N = 4         # chunks sent to the prompt after rerank

pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"])
claude = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY


# ---------- 1. chunking ----------
def split_text(text: str, size: int = CHUNK_SIZE, overlap: int = CHUNK_OVERLAP) -> list[str]:
    """Split on blank lines, pack paragraphs up to `size`, hard-split long paragraphs with overlap."""
    paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
    chunks, buf = [], ""
    for p in paragraphs:
        if len(p) > size:                       # long paragraph: flush buffer, then sliding window
            if buf:
                chunks.append(buf)
                buf = ""
            step = size - overlap
            chunks.extend(p[i:i + size] for i in range(0, len(p), step))
            continue
        if len(buf) + len(p) + 1 <= size:
            buf = f"{buf}\n{p}" if buf else p
        else:
            chunks.append(buf)
            # carry the tail of the previous chunk to reduce answers cut in half
            buf = (buf[-overlap:] + "\n" + p) if overlap else p
    if buf:
        chunks.append(buf)
    return chunks


def load_documents(root: Path) -> list[dict]:
    """Read .md / .txt files under root and build the documents to upsert."""
    docs = []
    for path in sorted(root.rglob("*")):
        if path.suffix.lower() not in {".md", ".txt"} or not path.is_file():
            continue
        rel = path.relative_to(root).as_posix()
        text = path.read_text(encoding="utf-8")
        # _id must be ASCII; hash the path so non-ASCII file names work
        file_key = hashlib.sha1(rel.encode("utf-8")).hexdigest()[:16]
        category = rel.split("/")[0] if "/" in rel else "general"
        for i, chunk in enumerate(split_text(text)):
            docs.append({
                "_id": f"{file_key}#{i}",
                "chunk_text": chunk,        # Pinecone embeds this field
                "source": rel,              # the rest is metadata, filterable automatically
                "category": category,
                "chunk_index": i,
            })
    return docs


# ---------- 2. index ----------
def ensure_index() -> None:
    if pc.indexes.exists(name=INDEX_NAME):
        return
    schema = (
        SchemaBuilder()
        .add_string_field(
            name="chunk_text",
            embed={"model": EMBED_MODEL, "dimension": EMBED_DIM},
        )
        .build()
    )
    print(f"Creating index {INDEX_NAME} ...")
    pc.indexes.create(
        name=INDEX_NAME,
        schema=schema,
        # managed = serverless. The free Starter plan only allows aws / us-east-1
        deployment={"deployment_type": "managed", "cloud": "aws", "region": "us-east-1"},
    )  # blocks until the index is ready by default


# ---------- 3. ingest ----------
def ingest(root: Path) -> None:
    ensure_index()
    index = pc.index(name=INDEX_NAME)
    docs = load_documents(root)
    if not docs:
        print("No .md / .txt files found")
        return

    # on re-ingest, delete each file's old chunks first so none are left behind
    for source in sorted({d["source"] for d in docs}):
        try:
            index.documents.delete(namespace=NAMESPACE, filter={"source": {"$eq": source}})
        except NotFoundError:
            pass  # namespace doesn't exist yet (first ingest)

    for i in range(0, len(docs), UPSERT_BATCH):
        batch = docs[i:i + UPSERT_BATCH]
        resp = index.documents.upsert(namespace=NAMESPACE, documents=batch)
        tokens = resp.usage.embed_total_tokens if resp.usage else "?"
        print(f"Upserted {i + len(batch)}/{len(docs)} chunks, embedding tokens this batch: {tokens}")

    time.sleep(5)  # indexing is asynchronous; wait a moment before querying
    print("Ingest complete")


# ---------- 4. retrieve + rerank ----------
def retrieve(question: str, category: str | None = None) -> list[dict]:
    index = pc.index(name=INDEX_NAME)
    resp = index.documents.search(
        namespace=NAMESPACE,
        top_k=RECALL_TOP_K,
        score_by=[{"type": "embed", "fields": ["chunk_text"], "query": question}],
        include_fields=["chunk_text", "source", "chunk_index"],
        filter={"category": {"$eq": category}} if category else None,
    )
    hits = [
        {
            "id": m.id,
            "score": m.score,
            "text": m.get("chunk_text", ""),
            "source": m.get("source", ""),
        }
        for m in resp.matches
    ]
    if not hits or not USE_RERANK:
        return hits[:FINAL_TOP_N]

    ranked = pc.inference.rerank(
        model=RERANK_MODEL,
        query=question,
        documents=[{"id": h["id"], "chunk_text": h["text"]} for h in hits],
        rank_fields=["chunk_text"],
        top_n=FINAL_TOP_N,
        return_documents=False,
        parameters={"truncate": "END"},
    )
    # r.index is the document's position in the list we passed in
    return [dict(hits[r.index], score=r.score) for r in ranked.data]


# ---------- 5. prompt Claude ----------
SYSTEM_PROMPT = """You are an internal knowledge-base assistant. Answer employees' questions using ONLY the material inside <documents>.
Rules:
1. If the answer isn't in the material, say "I couldn't find this in the knowledge base." Don't make anything up.
2. After each key statement, cite the source as [n], where n matches <document index="n">.
3. If sources conflict, list the conflict instead of picking one.
4. Answer concisely."""


def build_user_message(question: str, hits: list[dict]) -> str:
    blocks = "\n".join(
        f'<document index="{i}" source="{h["source"]}">\n{h["text"]}\n</document>'
        for i, h in enumerate(hits, start=1)
    )
    return f"<documents>\n{blocks}\n</documents>\n\nQuestion: {question}"


def answer(question: str, category: str | None = None) -> None:
    hits = retrieve(question, category)
    if not hits:
        print("Nothing was retrieved from the knowledge base.")
        return

    response = claude.beta.messages.create(
        model=CLAUDE_MODEL,
        max_tokens=16000,
        system=SYSTEM_PROMPT,
        messages=[{"role": "user", "content": build_user_message(question, hits)}],
        output_config={"effort": "medium"},
        # if a safety classifier refuses, the server retries on a recommended model (beta)
        fallbacks="default",
        betas=["server-side-fallback-2026-07-01"],
    )

    if response.stop_reason == "refusal":
        print("The model declined to answer.")
        return

    text = "".join(b.text for b in response.content if b.type == "text")
    print(text)
    print("\n---- Sources ----")
    for i, h in enumerate(hits, start=1):
        score = f"{h['score']:.4f}" if h["score"] is not None else "n/a"
        print(f"[{i}] {h['source']}  (score={score})")
    u = response.usage
    print(f"\n(model={response.model}, input_tokens={u.input_tokens}, output_tokens={u.output_tokens})")


# ---------- entry point ----------
def main() -> None:
    parser = argparse.ArgumentParser(description="Pinecone + Claude docs Q&A")
    sub = parser.add_subparsers(dest="cmd", required=True)
    p_ingest = sub.add_parser("ingest")
    p_ingest.add_argument("path", type=Path)
    p_ask = sub.add_parser("ask")
    p_ask.add_argument("question")
    p_ask.add_argument("--category", default=None)
    args = parser.parse_args()

    if args.cmd == "ingest":
        ingest(args.path)
    else:
        answer(args.question, args.category)


if __name__ == "__main__":
    try:
        main()
    except anthropic.APIStatusError as e:
        sys.exit(f"Claude API error {e.status_code}: {e.message}")

Running it

Prepare a folder whose top-level subfolders act as categories:

docs/
├── hr/
│   ├── annual-leave.md
│   └── onboarding.md
└── finance/
    └── expenses.md

python rag_bot.py ingest ./docs
python rag_bot.py ask "Do I get annual leave in my first year?"
python rag_bot.py ask "What's the hotel limit on business trips?" --category finance
USE_RERANK=0 python rag_bot.py ask "How far ahead do I need to request leave?"   # no rerank quota used

The output shows the answer, the source file behind each [n] with its score, and the token usage for the call so you can check it against your bill.

10. Estimating cost

These are list-price estimates, not a measured bill; your console usage is the source of truth.

Pinecone (as of November 2026, Pricing):

An internal knowledge base of a few hundred policy docs usually fits within Starter's 2 GB and 5M embedding tokens a month. The first limit you'll hit is usually 500 reranks a month. Either disable rerank on Starter, call it only when recall scores are too close to separate, or upgrade. Read/write unit accounting is detailed in Understanding cost.

Claude (as of November 2026, Pricing): claude-sonnet-5-5 is $2 per million input tokens, $10 per million output tokens, and $0.10 per million cache-hit tokens.

Rough math for one question with these settings: four chunks plus system prompt and question is a few thousand input tokens; the answer is a few hundred. Input cost is well under a cent; output cost depends on answer length, and thinking tokens are billed as output, which is why effort is worth tuning. For real numbers, multiply the input_tokens / output_tokens the script prints by the unit prices.

Ways to save:

• Bigger top_k and FINAL_TOP_N aren't automatically better: 10 chunks instead of 4 is more than double the input tokens.
• The system prompt never changes, so at volume consider Claude prompt caching. There's a minimum cacheable length, so a very short system prompt won't benefit; see the docs.
• Use Pinecone namespaces per department or tenant; queries only scan their namespace, which also cuts read units.

11. Common pitfalls

1. Dimension mismatch

If you generate vectors yourself (options B/C), the index dimension must match the model output. Writing with model A and querying with model B gives essentially random results even if the dimensions happen to match. Using an embed field with embed scoring avoids this structurally, because Pinecone uses the model fixed at creation for both. Remember: the schema and embedding config can't change after creation; a new model means a new index and a full re-ingest.

With option B, the quickstart pattern is pc.inference.embed(model=..., inputs=[...], parameters={"input_type": "passage"}), written into a field declared with add_dense_vector_field, and input_type="query" at search time. Omitting or swapping input_type won't error, but retrieval quality drops.

2. Wrong namespace: "successful" query, empty results

If the write and query namespaces differ, you get no error, just empty results. The script keeps the namespace in one env var. Namespaces are good for multi-tenant isolation (one per department or customer), but Starter caps an index at 100 namespaces, so many tenants need a higher plan or metadata filtering instead.

3. Metadata filtering rules

• Any field not declared in the schema is metadata and is indexed for filtering on write, supporting $eq, $in, $gte and friends.
• Scoring fields declared in the schema (like chunk_text) can't use comparison operators like $eq (400 error); only text-match operators like $match_phrase / $match_all / $match_any.
• Metadata is capped at 40 KB per document; don't stuff the full source text into it.
• Keep each field's type consistent, e.g. chunk_index always an integer.

Metadata filters are not access control. If finance policies should only be visible to finance staff, the server must add the filter based on the logged-in user; never trust a category sent from the client. For stricter isolation, use separate namespaces or even separate indexes.

4. Chunk granularity

• Too large: input beyond the embedding model's limit is silently truncated (default truncate=END), so the tail never makes it into the vector.
• Too small: the search hits, but the model gets half a sentence.
• For policy docs, prefix each chunk with the doc or section title, e.g. "Travel Expense Policy, Section 3, Lodging: ...". It helps retrieval noticeably and costs almost nothing.
• Tables and scanned PDFs need text extraction first; that's outside Pinecone's job.

5. Document IDs and batch size

• IDs must be ASCII; convert non-ASCII file names.
• Indexes with an embed field: max 96 documents per upsert.
• Indexes with an embed field don't support bulk import from object storage; use upsert / update only.

6. Not found right after writing

Indexing is asynchronous, so there's a short eventual-consistency window. A few seconds of sleep is enough in an ingest script; if your app needs read-after-write, handle it with a retry or a UI hint.

7. Mixing old and new SDK styles

8. Claude-side pitfalls

• Passing temperature returns 400 (SDK 2.x / API version 2026-10-21).
• Don't prefill an assistant message at the end of messages to steer formatting; current models don't support prefill and return 400. Use output_config.format for structured output.
• Filter text on type == "text".
• Handle stop_reason == "refusal" and truncation (stop_reason == "max_tokens").

9. Data residency and compliance

Check these before rolling this out inside a company:

• Starter indexes live only in AWS US East. Other regions (including Singapore ap-southeast-1 and Europe) need a paid plan (Cloud regions).
• Hosted embedding and rerank models have their own processing locations; some run in global, possibly outside the US (Processing locations).
• Pinecone's site says GDPR and ISO 27001 coverage doesn't apply to Starter, and HIPAA needs a Standard add-on or Enterprise. For stricter needs, look at Enterprise BYOC (deployed in your own cloud account).
• Calls to the Claude API also send your text to Anthropic; check the supported-countries list and your company's data policies.

If your docs contain personal data or sensitive business information, get legal and security review first. If data can't leave your environment, this managed stack isn't the right fit; look at self-hostable options like pgvector or Milvus with a model you can run or contract for in-region.

12. Summary

1. Chunk: pack paragraphs, hard-split long ones, keep overlap; attach source and category metadata.
2. Embed + write: declare embed on a string field of a document index, write plain text, let Pinecone embed it with llama-text-embed-v2; at most 96 per batch.
3. Retrieve: documents.search with embed scoring recalls 12, optionally filtered by metadata.
4. Rerank: pc.inference.rerank with bge-reranker-v2-m3 keeps the best 4.
5. Generate: numbered chunks go into <documents> tags; Claude Sonnet 5.5 answers only from them, with citations.

The upside is very little code and no models or databases to operate, which makes it a good way to test whether "Q&A over company docs" is worth building. The downsides are just as clear: Pinecone is a closed managed service and your data lives in its cloud; Starter's rerank and region limits are tight; and costs grow linearly with data and query volume, so reconcile your bill regularly.

Where to go next:

• Prefix chunks with titles, or chunk by Markdown heading.
• Hybrid search: enable full-text search on the same field and merge keyword and vector results with RRF; it helps with exact terms like product codes or policy numbers.
• Build a small eval set from real questions, track "answer correct" and "citation correct", then tune CHUNK_SIZE, top_k and effort.
• Hook it into Slack or Teams as a bot, with server-side filtering by user identity.

Pinecone and Anthropic iterate quickly; check versions before running, and where this article and the official docs disagree, trust the docs.

References

• Pinecone, Ingest your own files: https://docs.pinecone.io/guides/get-started/quickstart/ingest-files.md
• Pinecone, Create an index: https://docs.pinecone.io/guides/index-data/create-an-index.md
• Pinecone, Integrated embedding: https://docs.pinecone.io/guides/search/integrated-embedding.md
• Pinecone, Index configuration (regions, embedding models): https://docs.pinecone.io/guides/index-data/index-configuration.md
• Pinecone, Rerank results: https://docs.pinecone.io/guides/search/rerank-results.md
• Pinecone, Rate limits: https://docs.pinecone.io/reference/api/database-limits/rate-limits.md
• Pinecone, Object limits: https://docs.pinecone.io/reference/api/database-limits/object-limits.md
• Pinecone, Understanding cost: https://docs.pinecone.io/guides/manage-cost/understanding-cost.md
• Pinecone, Processing locations: https://docs.pinecone.io/guides/index-data/processing-locations.md
• Pinecone, 2026 release notes: https://docs.pinecone.io/release-notes/2026.md
• Pinecone, Pricing: https://www.pinecone.io/pricing/
• Pinecone Python SDK (PyPI): https://pypi.org/project/pinecone/
• Claude, Pricing: https://platform.claude.com/docs/en/about-claude/pricing
• Claude, Refusals and fallback: https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback
• Claude, Embeddings: https://platform.claude.com/docs/en/build-with-claude/embeddings
• Anthropic Python SDK: https://github.com/anthropics/anthropic-sdk-python
• Anthropic supported countries: https://www.anthropic.com/supported-countries