Agentsop Idempotent Ingestion

Idempotent Ingestion · Re-Ingest-Correctness SOP

> Third-person operating model for a coder agent that owns ingestion

> correctness across repeated runs. The audience is the LLM agent writing or

> reviewing the pipeline code — not the end user.

> One sentence: *The interesting run is the second one. A correct pipeline

> hashes each document, looks the hash up in a docstore, and inserts / updates

> / skips accordingly — so re-running over unchanged docs changes nothing.*

This skill is an ENHANCE overlay over [[llamaindex]]. The base skill names

IngestionPipeline with docstore + UPSERTS_AND_DELETE as a hardening

must-have (OP-08, anti-pattern A10) but stops at "use it". This overlay

surfaces the *correctness contract* — what idempotency actually means, how the

hash decides the action, and how it generalises beyond LlamaIndex.

1. 何时激活 (Activation Rules)

Activate this skill whenever any of the following holds:

• The codebase contains an ingestion call (VectorStoreIndex.from_documents,

pipeline.run(documents=...), vectorstore.add_documents, a custom embed

+ upsert loop) and that call will execute more than once over a corpus

that can change between runs.

• The user mentions any of: re-ingest, re-index, scheduled / nightly index

refresh, incremental document updates, docstore, doc hash, upsert,

RecordManager, "keep the index in sync with the source".

• A bug report says "I see the same answer chunk twice", "retrieval returns

duplicates", "deleted a file but it still shows up in answers", "the index

keeps growing every night even though nothing changed".

• PR review: new ingestion code that builds the index with no docstore / no

hash-keyed dedup and the corpus is live (LlamaIndex base skill A10).

• A production RAG is about to ship and the ingestion side has only ever been

tested as a cold first run.

Do not activate when:

• The index is built once from a static corpus and never refreshed — there

is no second run to make idempotent.

• The corpus is tiny and rebuilding from scratch each run is *measurably*

cheaper than maintaining a docstore (rare; verify, don't assume).

• The task is purely query-time (retrieval, rerank, synthesis) with no

ingestion path in scope.

2. 核心心智模型 (Core Mental Model)

Three principles. Violating any one means a second run can duplicate or corrupt

the index — regardless of how good the retrieval stack downstream is.

Principle 1 — Ingestion must be idempotent; the hash is the decision

> Re-running the pipeline over an unchanged document is a no-op. Run it

> once, run it a thousand times — same index. The mechanism: each document is

> reduced to a stable content hash (and a stable doc id); that hash is

> looked up in a docstore (a key→hash record store separate from the vector

> store); the lookup result decides the action.

            doc_id seen?           hash changed?
                │                       │
   ┌────────────┴───────────┐    ┌──────┴──────┐
   no                       yes  no            yes
   │                         │   │              │
INSERT (embed + upsert)      │  SKIP        UPDATE
                             └──────────────►(delete old chunks,
                                              re-embed, upsert)

The hash is the whole game. Without it, the pipeline cannot tell "I have seen

this exact content" from "this is new" — so it re-embeds and re-adds

everything, and the vector store accumulates duplicates. (LlamaIndex base skill

failure #3 / A10; developers.llamaindex.ai/.../loading/ingestion_pipeline/.)

Principle 2 — The docstore is a separate ledger, not the vector store

The vector store holds embeddings keyed by node id. The docstore holds, per

document, {doc_id → content_hash} (and the node ids derived from it). They are

two stores with two jobs:

| Store | Holds | Answers |

|---|---|---|

| Docstore | doc_id → hash, doc→node mapping | "have I seen this content before?" |

| Vector store | node_id → embedding + metadata | "what is semantically near this query?" |

The dedup decision happens against the docstore, *before* any embedding

call — so an unchanged corpus costs zero embedding tokens on re-run. If the

docstore is ephemeral (in-memory, lost on restart), every cold start looks like

a first run and re-embeds the world. **The docstore must be persisted to the

same durability tier as the vector store** (SimpleDocumentStore.persist(),

Redis, MongoDB, Postgres). A docstore that doesn't survive the process is not a

docstore.

Principle 3 — Deletes don't propagate for free

Insert and update are driven by *documents that are present*. A document that

disappeared from the source emits no event — so a naive upsert pipeline

never learns it should remove the orphaned chunks. Stale content keeps getting

retrieved long after the source file is gone. Propagating deletes requires the

pipeline to compare the full set of doc_ids seen this run against the set in

the docstore, and purge the difference. In LlamaIndex this is the _AND_DELETE

half of DocstoreStrategy.UPSERTS_AND_DELETE; in LangChain it is

cleanup="full" / cleanup="incremental". Choosing the upsert-only strategy

silently accepts stale ghosts — sometimes correct (append-only corpus), usually

not.

3. SOP 工作流 (Agentic Protocol)

Five stages. Each gates the next. Stop and remediate at the first failure. The

gate that matters most is Stage 4 — *prove the second run is a no-op*.

Stage 0 — Frame the re-run

Before code, answer:

• Run cadence: how does the pipeline get re-triggered? (cron / webhook /

CI / manual). If the honest answer is "never, it's one-shot" → this skill is

over-engineering; stop.

• Change shape: do source docs get *added* only, or also *edited* and

*deleted*? This decides upsert-only vs upsert-and-delete (Principle 3).

• Doc identity: what is the stable doc_id? It must survive across runs —

a file path, a CMS id, a URL. Not a random uuid generated at load time

(that makes every run look new). LlamaIndex derives a hash from content +

doc_id; if doc_id is unstable, idempotency breaks even with a docstore.

• Durability tier: where does the docstore live so it survives restarts?

Stage 1 — Attach a persisted docstore

from llama_index.core.ingestion import IngestionPipeline, DocstoreStrategy
from llama_index.core.storage.docstore import SimpleDocumentStore
from llama_index.core.node_parser import SentenceSplitter
from llama_index.embeddings.openai import OpenAIEmbedding

pipeline = IngestionPipeline(
    transformations=[
        SentenceSplitter(chunk_size=1024, chunk_overlap=20),
        OpenAIEmbedding(model="text-embedding-3-small"),
    ],
    docstore=SimpleDocumentStore(),          # the dedup ledger
    vector_store=vector_store,               # the embedding store
    docstore_strategy=DocstoreStrategy.UPSERTS_AND_DELETE,
)

For production, swap SimpleDocumentStore for a server-backed one

(RedisDocumentStore, MongoDocumentStore, PostgresDocumentStore) so the

ledger is shared across workers and survives restarts. A SimpleDocumentStore

that is never .persist()-ed is the #1 cause of "it re-embeds everything every

night".

Stage 2 — Pin stable doc ids and let the pipeline hash

docs = SimpleDirectoryReader("./data", filename_as_id=True).load_data()
# filename_as_id=True → doc.id_ is the file path → stable across runs.
# LlamaIndex then computes the content hash internally per node.
pipeline.run(documents=docs)

Hard rules:

• doc_id comes from a stable external identity (path / CMS id / URL),

never a fresh uuid per load. Set filename_as_id=True or assign doc.id_

explicitly.

• Do not mutate document text in a non-deterministic transformation before

the hash is taken (e.g. injecting datetime.now() into metadata) — that makes

the hash change every run and defeats the skip path. Keep volatile metadata

out of the hashed payload.

• The embedding model is part of the index identity: changing it requires a

full re-embed (LlamaIndex base skill A2), not an incremental upsert.

Stage 3 — Choose the upsert strategy deliberately

| Strategy | Re-run behaviour | Use when |

|---|---|---|

| UPSERTS | new+changed docs upserted; deletes not propagated | append-mostly corpus; deletes handled out-of-band |

| UPSERTS_AND_DELETE | new+changed upserted; vanished docs purged | the corpus is a *mirror* of a source that can shrink |

| DUPLICATES_ONLY | dedup identical docs *within one run*; no cross-run update | one-shot batch that may contain dupes; no live sync |

Default for any system that mirrors a mutable source: UPSERTS_AND_DELETE.

DUPLICATES_ONLY is not a re-ingest strategy — it only de-dupes within a

single .run(); a second .run() with the same docs will *not* skip them.

Stage 4 — Prove the second run is a no-op (the gate)

Before merging, the PR must include — and CI must run — a test that runs the

pipeline twice and asserts the second run did nothing observable:

def test_reingest_is_idempotent(pipeline, docs, vector_store):
    n1 = len(pipeline.run(documents=docs))      # cold run
    count_after_first = vector_store.count()

    n2 = len(pipeline.run(documents=docs))       # identical re-run
    count_after_second = vector_store.count()

    assert n2 == 0,                       "re-run re-processed unchanged docs"
    assert count_after_second == count_after_first, "re-run duplicated vectors"

pipeline.run() returns the list of nodes it actually *processed*; on a

clean idempotent re-run over unchanged input that list is empty. If it

isn't, the docstore is missing, ephemeral, or the doc ids / hashes are

unstable — go back to Stage 1-2. Then add the edit case (change one doc → exactly

that doc's chunks replaced) and, for UPSERTS_AND_DELETE, the delete case

(remove one doc → its chunks gone, others untouched). See OP-06 / OP-07.

Stage 5 — Schedule, persist, observe

• Persist the docstore *after every run* (storage_context.persist() /

server-backed store auto-persists) — otherwise the next cold start re-embeds

everything.

• Run on a schedule, not by hand (LlamaIndex base skill OP-08): cron / Airflow /

serverless cron. The whole point of idempotency is that a missed-or-doubled

trigger is harmless.

• Log per run: `{docs_in, inserted, updated, skipped, deleted,

embedding_calls}. On a steady-state corpus, inserted+updated+deleted` should

trend to ~0 and skipped ≈ docs_in. A run that re-embeds everything every

night is the loud signal that idempotency is broken.

4. 操作模型 (Operation Models)

Format: Trigger / Action / Output / Evidence. Machine-readable in

intermediate/operation_candidates.json.

OP-01 AttachDocstore

• Trigger: any IngestionPipeline / ingestion path that will run more than

once with no docstore wired.

• Action: pass docstore= (persisted: Redis/Mongo/Postgres in prod,

SimpleDocumentStore only for tests) plus docstore_strategy=. The docstore

is the dedup ledger; without it dedup is structurally impossible.

• Output: pipeline that can answer "have I seen this content?" before

embedding.

• Evidence: developers.llamaindex.ai/python/framework/module_guides/loading/ingestion_pipeline/; [[llamaindex]] OP-08 / A10.

OP-02 PinStableDocId

• Trigger: docs loaded with random / per-run ids; "re-run re-adds

everything" symptom.

• Action: set filename_as_id=True or assign doc.id_ from a stable

external identity (path / CMS id / URL). Never a fresh uuid per load.

• Output: the same source document hashes to the same key across runs → the

skip path can fire.

• Evidence: LlamaIndex Document.doc_id / SimpleDirectoryReader(filename_as_id=True) docs.

OP-03 HashBeforeEmbed

• Trigger: designing or reviewing the dedup decision point.
• Action: ensure the content hash is computed and looked up in the docstore

*before* the embedding transformation runs, so unchanged docs cost zero

embedding tokens. In LlamaIndex this is automatic given a docstore; in a

manual pipeline, hash first, embed only on miss/change.

• Output: re-run cost on a steady corpus ≈ 0 embedding calls.
• Evidence: IngestionPipeline dedup section; LangChain indexing API

("avoids re-writing unchanged content, avoids re-computing embeddings").

OP-04 PickUpsertStrategy

• Trigger: choosing pipeline behaviour for the corpus's change shape.
• Action: UPSERTS (append-mostly), UPSERTS_AND_DELETE (mirror of a

shrinkable source — the production default), DUPLICATES_ONLY (in-run dedup

only, not cross-run). Decide from Stage 0 change shape, not by default.

• Output: documented strategy + one-line rationale.
• Evidence: DocstoreStrategy enum; LlamaIndex ingestion docs.

OP-05 PropagateDeletes

• Trigger: source docs can be deleted and stale chunks must not be

retrieved.

• Action: use UPSERTS_AND_DELETE; the pipeline diffs the doc_ids seen this

run against the docstore and purges orphans. (LangChain: cleanup="full" for

a full snapshot run, cleanup="incremental" for streamed batches.)

• Output: deleted source docs → their vectors removed; retrieval stays in

sync.

• Evidence: DocstoreStrategy.UPSERTS_AND_DELETE; LangChain `index(...,

cleanup=...)` docs.

OP-06 ReingestNoOpTest

• Trigger: any ingestion code change merges.
• Action: run the pipeline twice over identical input; assert the second

run() returns 0 processed nodes and the vector count is unchanged (Stage 4).

• Output: regression gate proving idempotency is wired through.
• Evidence: pipeline.run() returns processed nodes; direct application of

Principle 1.

OP-07 IncrementalUpdateTest

• Trigger: corpus supports edits and/or deletes.
• Action: edit one doc → assert exactly that doc's old chunks are gone and

new ones present, others untouched. Delete one doc (under

UPSERTS_AND_DELETE) → assert its chunks gone, neighbours intact.

• Output: proof that update/delete are surgical, not full-rebuild.
• Evidence: hash-change → update path; delete path of

UPSERTS_AND_DELETE.

OP-08 PersistDocstore

• Trigger: SimpleDocumentStore used, or docstore not persisted after run.
• Action: storage_context.persist(persist_dir=...) after each run, or use

a server-backed docstore that auto-persists and is shared across workers.

• Output: cold start reads the ledger → does not re-embed the world.
• Evidence: LlamaIndex storage / persistence docs;

Redis/Mongo/Postgres docstore integrations.

OP-09 ManualHashUpsert

• Trigger: no framework available (raw vector-store SDK, custom loader).
• Action: maintain a {doc_id → sha256(content)} table next to the vector

store; on each doc compute the hash, compare, and INSERT / UPDATE

(delete-then-add chunks) / SKIP; track seen ids for delete propagation.

• Output: framework-free idempotent ingestion with the same contract.
• Evidence: this is what IngestionPipeline / RecordManager automate; see

R2 cross-framework table.

5. 困境决策案例 (Dilemma Cases)

Dilemma 1 — A document changed slightly: re-embed all, or diff?

困境: A 200-page handbook had one paragraph edited. The naive idempotent

pipeline hashes at the document level → the whole doc's hash changes → all

its chunks are deleted and re-embedded. Correct for safety, but on a large doc a

one-line edit triggers a full re-embed of that document. Should the pipeline

diff at the chunk level instead?

约束:

• LlamaIndex's docstore dedup keys on the document hash by default — any

change re-processes the whole document's nodes.

• Chunk-level diffing requires stable chunk identity across edits, but

SentenceSplitter re-segments the whole doc when text shifts, so chunk

boundaries (and ids) move even for unedited text downstream of the edit.

• Re-embedding one document is usually cheap; re-embedding the *whole corpus*

every run (the failure this skill prevents) is the expensive thing.

决策步骤:

• Default to document-level hashing — it is correct and simple. A

per-document re-embed on edit is acceptable for most corpora.

• If single documents are huge *and* edited frequently, **split the source into

smaller logical documents** (per section / per page) so the hash granularity

matches the edit granularity — each becomes its own doc_id.

• Only build true chunk-level content-addressed dedup (hash each chunk, stable

chunk id) if profiling proves per-document re-embed is the real bottleneck.

It is rarely worth the complexity.

结果: Move the granularity by splitting documents, not by hand-rolling

chunk-diffing. Document-level hashing stays the default; the fix is upstream in

how the source is segmented into docs.

可提取的操作: OP-02, OP-03, OP-04.

Dilemma 2 — Source documents were deleted: purge from the index, or keep?

困境: 50 files were removed from the source folder. The upsert-only pipeline

leaves their chunks in the vector store forever — they keep surfacing in

retrieval and the LLM cites documents that no longer exist. Switch to

UPSERTS_AND_DELETE? But what if a run sees a *partial* corpus (a flaky reader

returned only half the files) — would it then delete everything legitimately

missing-this-run?

约束:

• UPSERTS_AND_DELETE purges any doc_id in the docstore that was not seen

in the current run. If the run's input is incomplete, that purge is

catastrophic — it deletes live content.

• Therefore delete-propagation is only safe when the run input is a **complete

snapshot** of the source (LangChain calls this cleanup="full"); for streamed

/ partial batches you need cleanup="incremental" (per-batch, scoped by

source).

• Stale chunks are a *correctness* bug (wrong citations); over-aggressive delete

is an *availability* bug (lost content). Both are bad.

决策步骤:

• Confirm the run input is a full snapshot. If the reader can return a

partial set on failure, gate the delete: abort the run (don't purge) if the

seen-doc count drops more than a sane threshold vs the previous run.

• Snapshot + complete → UPSERTS_AND_DELETE (LlamaIndex) / cleanup="full"

(LangChain). Streamed batches scoped by source → cleanup="incremental".

• For regulated corpora, prefer soft delete (deleted=true metadata +

query-time filter) over hard delete, so erasure is auditable and reversible.

结果: Purge — but only behind a "this is a complete snapshot" guard.

Idempotency for deletes is UPSERTS_AND_DELETE *plus* a partial-input circuit

breaker, never a blind diff-and-delete.

可提取的操作: OP-05, OP-07, OP-08.

6. 反模式与边界 (Anti-patterns & Boundaries)

Top anti-patterns (instant red flags in code review)

| # | Anti-pattern | Why it's wrong | Correct move |

|---|---|---|---|

| A1 | VectorStoreIndex.from_documents(docs) re-run every cron tick | No docstore → re-embeds + re-adds everything; vector store grows without bound; duplicates pollute top-k | IngestionPipeline(docstore=..., docstore_strategy=UPSERTS_AND_DELETE) (OP-01) |

| A2 | IngestionPipeline(...) with no docstore= | The dedup decision has nowhere to look up the hash — same as no idempotency | Pass a persisted docstore ([[llamaindex]] A10) |

| A3 | SimpleDocumentStore() never .persist()-ed | Ledger lost on restart; every cold start looks like a first run | Persist after each run, or use Redis/Mongo/Postgres docstore (OP-08) |

| A4 | Random / per-load doc_id (fresh uuid each run) | Same content hashes to a new key each run → nothing ever skips | Stable external id: filename_as_id=True / explicit doc.id_ (OP-02) |

| A5 | Volatile metadata (timestamps, run id) in the hashed payload | Hash changes every run even for unchanged content → constant re-embed | Keep volatile metadata out of the hashed body |

| A6 | DUPLICATES_ONLY treated as a re-ingest strategy | Only dedupes *within one run*; a second run does not skip | Use UPSERTS / UPSERTS_AND_DELETE for cross-run (OP-04) |

| A7 | Upsert-only on a corpus whose source files get deleted | Stale chunks linger; LLM cites deleted docs | UPSERTS_AND_DELETE behind a snapshot guard (OP-05, Dilemma 2) |

| A8 | Blind UPSERTS_AND_DELETE on a possibly-partial run input | A flaky reader returning half the files purges live content | Gate delete on "complete snapshot"; circuit-break on big drops |

| A9 | No twice-run test in CI | First duplicate / ghost found by a user in production | OP-06 mandatory on every PR |

| A10 | Swapping embedding model and expecting incremental upsert to "fix" old vectors | Old vectors were embedded by the old model; mixed space breaks retrieval | Full re-embed; tag index with model name+version ([[llamaindex]] A2) |

Boundaries — when this skill is not the right move

• B1 One-shot index, never refreshed → no second run; skip.
• B2 Corpus tiny enough that full rebuild each run is *measurably* cheaper

than docstore maintenance → rebuild is fine; but profile before deciding.

• B3 Append-only event log that never edits or deletes → UPSERTS (or even

plain add) is sufficient; delete propagation is dead weight.

• B4 The "duplicate results" bug is actually a chunk-overlap or

retriever top_k issue, not a re-ingest issue → diagnose first; don't add a

docstore to a problem it doesn't solve.

PR review smells

• from_documents(...) inside anything that runs on a schedule.
• IngestionPipeline(...) with no docstore= keyword.
• SimpleDocumentStore() and no persist( anywhere in the file.
• doc.id_ = str(uuid4()) / loader without filename_as_id.
• datetime.now() / time.time() written into document metadata *before*

ingestion.

• docstore_strategy=DocstoreStrategy.DUPLICATES_ONLY on a live-sync pipeline.
• A re-ingest function with no test that calls it twice.

7. 跨框架对照 (Cross-Framework Reference Table)

The same "hash → docstore → insert/update/skip" contract across the common

stacks. Compact below; full runnable code in references/R2-cross-framework.md.

Verified against current docs (May 2026).

• LlamaIndex — `IngestionPipeline(docstore=..., vector_store=...,

docstore_strategy=DocstoreStrategy.UPSERTS_AND_DELETE)`. Hash is on the

document (content + doc_id); pipeline.run(documents=docs) returns only the

nodes it *processed* — empty list = clean no-op. Persist the docstore

(Redis/Mongo/Postgres in prod) to keep idempotency across restarts.

• LangChain — `index(docs, SQLRecordManager(...), vectorstore,

cleanup="full", source_id_key="source"). The RecordManager` is the

docstore-equivalent (a hash per source_id in a durable store). Returns

{num_added, num_updated, num_skipped, num_deleted}; on an unchanged re-run

num_skipped == len(docs) and the rest are 0. cleanup="full" propagates

deletes for a complete snapshot, "incremental" for source-scoped batches.

• Manual — keep a persisted {doc_id: sha256(text)} table; per doc:

prev == h → SKIP, prev set → delete-then-add (UPDATE), else INSERT; after

the loop purge stored_keys − seen (DELETE); persist(). This is exactly

what the two frameworks automate.

7.x Side-by-side: "what happens on the second run?"

| Stack | Mechanism | Unchanged re-run | Deletes propagate? |

|---|---|---|---|

| LlamaIndex IngestionPipeline + docstore | document hash in docstore | run() returns [] | only with UPSERTS_AND_DELETE |

| LangChain index() + RecordManager | hash per source_id in SQL/record store | num_skipped == len(docs) | only with cleanup="full"/"incremental" |

| Manual hash table | {doc_id: sha256} table | inner loop all continue | only if you diff seen vs stored keys |

| from_documents / raw add (no ledger) | none | re-embeds + re-adds → duplicates | never |

The bottom row is the anti-pattern (A1). Every correct row has the same shape:

a persisted hash ledger separate from the vector store, consulted before

embedding, with an explicit delete-propagation switch.

References

Primary docs (cited inline above)

• LlamaIndex — *Ingestion Pipeline* (docstore dedup, DocstoreStrategy):

https://developers.llamaindex.ai/python/framework/module_guides/loading/ingestion_pipeline/

• LlamaIndex — *Document Management* (docstore + UPSERTS):

https://developers.llamaindex.ai/python/framework/module_guides/loading/ingestion_pipeline/document_management/

• LlamaIndex — *Storage / persisting* (docstore durability):

https://developers.llamaindex.ai/python/framework/module_guides/storing/

• LangChain — *Indexing* (RecordManager, cleanup):

https://python.langchain.com/docs/how_to/indexing/

Companion files

• references/R1-source-evidence.md — claims traced to base [[llamaindex]] skill + primary docs.
• references/R2-cross-framework.md — full LlamaIndex / LangChain / manual code with delete-propagation variants.
• intermediate/operation_candidates.json — OP-01..09 in machine-readable Trigger / Action / Output / Evidence form.

Base skill (overlay relationship)

• [[llamaindex]] — base RAG SOP. This overlay enhances OP-08

(IngestionWithDocstore) and anti-pattern A10 with the full

re-ingest-correctness contract.