embedflow 0.2.0__tar.gz → 0.4.0__tar.gz

{embedflow-0.2.0 → embedflow-0.4.0}/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## v0.4.0 — Milvus backend
+
+- Added a read-only Milvus backend for existing dense `FLOAT_VECTOR`
+  collections.
+- Added collection/database/partition selection, HNSW/IVF search parameters,
+  metric and schema auditing, and Milvus-backed document text resolution.
+- Added deterministic standalone Docker fixtures, examples, and optional
+  `pymilvus` packaging.
+
+## v0.3.0 — Pinecone backend
+
+- Added a read-only Pinecone backend for existing dense indexes.
+- Added host targeting, index-name resolution, namespace-aware retrieval, and
+  metadata or external-document text resolution.
+- Added Pinecone status/audit integration, optional dependency packaging, and
+  unit/integration smoke fixtures.
+
 ## v0.2.0 — pgvector backend
 
 - Added a read-only pgvector backend for existing PostgreSQL vector tables.

{embedflow-0.2.0 → embedflow-0.4.0}/CITATION.cff

@@ -2,7 +2,7 @@ cff-version: 1.2.0
 title: "EmbedFlow: Upgrading Legacy Embeddings Without Full Upfront Re-Embedding"
 message: "If EmbedFlow contributes to your work, please cite this software release."
 type: software
-version: 0.2.0
+version: 0.4.0
 date-released: 2026-09-06
 repository-code: "https://github.com/arnsri33/embedflow"
 url: "https://github.com/arnsri33/embedflow"

{embedflow-0.2.0 → embedflow-0.4.0}/CONTRIBUTING.md

@@ -11,7 +11,7 @@ python -m pip install -e '.[dev]'
 ```
 
 Optional integrations can be installed with `.[faiss]`, `.[qdrant]`,
-`.[models]`, or `.[dashboard]`.
+`.[pgvector]`, `.[pinecone]`, `.[models]`, or `.[dashboard]`.
 
 ## Checks before opening a pull request
 

{embedflow-0.2.0 → embedflow-0.4.0}/MANIFEST.in

@@ -1,6 +1,6 @@
 include README.md README_PYPI.md LICENSE CHANGELOG.md CONTRIBUTING.md SECURITY.md CITATION.cff requirements.txt requirements-dev.txt
 recursive-include docs *
-recursive-include examples *.md *.yaml *.json *.jsonl *.py *.sql *.sh
+recursive-include examples *.md *.yaml *.yaml.example *.json *.jsonl *.py *.sql *.sh
 recursive-include scripts *.sh *.py
 recursive-include frozen *.md *.sha256
 prune .github

{embedflow-0.2.0 → embedflow-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: embedflow
-Version: 0.2.0
+Version: 0.4.0
 Summary: Progressive embedding-model migration over existing vector indexes.
 Author: Arnav Srivastav
 License-Expression: AGPL-3.0-only
@@ -8,7 +8,7 @@ Project-URL: Homepage, https://embedflow.org
 Project-URL: Repository, https://github.com/arnsri33/embedflow
 Project-URL: Documentation, https://github.com/arnsri33/embedflow#readme
 Project-URL: Issues, https://github.com/arnsri33/embedflow/issues
-Keywords: embeddings,vector-search,rag,information-retrieval,faiss,qdrant,pgvector
+Keywords: embeddings,vector-search,rag,information-retrieval,faiss,qdrant,pgvector,pinecone,milvus
 Classifier: Development Status :: 3 - Alpha
 Classifier: Intended Audience :: Developers
 Classifier: Intended Audience :: Science/Research
@@ -29,6 +29,10 @@ Provides-Extra: qdrant
 Requires-Dist: qdrant-client>=1.9; extra == "qdrant"
 Provides-Extra: pgvector
 Requires-Dist: psycopg[binary]>=3.2; extra == "pgvector"
+Provides-Extra: pinecone
+Requires-Dist: pinecone>=6.0; extra == "pinecone"
+Provides-Extra: milvus
+Requires-Dist: pymilvus<4,>=2.5.5; extra == "milvus"
 Provides-Extra: models
 Requires-Dist: huggingface-hub<1.0,>=0.34; extra == "models"
 Requires-Dist: transformers<5.0,>=4.45; extra == "models"
@@ -42,6 +46,8 @@ Provides-Extra: all
 Requires-Dist: faiss-cpu>=1.8.0; extra == "all"
 Requires-Dist: qdrant-client>=1.9; extra == "all"
 Requires-Dist: psycopg[binary]>=3.2; extra == "all"
+Requires-Dist: pinecone>=6.0; extra == "all"
+Requires-Dist: pymilvus<4,>=2.5.5; extra == "all"
 Requires-Dist: huggingface-hub<1.0,>=0.34; extra == "all"
 Requires-Dist: transformers<5.0,>=4.45; extra == "all"
 Requires-Dist: sentence-transformers>=3.0; extra == "all"
@@ -62,7 +68,7 @@ Dynamic: license-file
 EmbedFlow lets a new embedding model serve over candidates from an existing
 vector index while target document vectors are materialized progressively. It
 supports migration analysis, persistent caching, background work, FAISS,
-Qdrant, pgvector, a CLI, and FastAPI.
+Qdrant, pgvector, Pinecone, Milvus, a CLI, and FastAPI.
 
 The full project README and architecture diagram are on
 <https://github.com/arnsri33/embedflow>.
@@ -85,6 +91,18 @@ For an existing PostgreSQL/pgvector table:
 python -m pip install "embedflow[pgvector]"
 ```
 
+For an existing Pinecone dense index:
+
+```bash
+python -m pip install "embedflow[pinecone]"
+```
+
+For an existing Milvus collection:
+
+```bash
+python -m pip install "embedflow[milvus]"
+```
+
 Qdrant and model-runtime extras are documented in the
 [installation guide](https://github.com/arnsri33/embedflow/blob/main/docs/installation.md).
 For model-backed analysis, install `embedflow[faiss,models,dashboard]`.
@@ -186,10 +204,14 @@ for definitions and reproduction details.
 | FAISS | Supported |
 | Qdrant | Supported |
 | pgvector | Supported |
+| Pinecone | Supported |
+| Milvus | Supported |
 
 See the [FAISS guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/faiss.md),
 [Qdrant guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/qdrant.md),
-and [pgvector guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pgvector.md).
+and [pgvector guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pgvector.md),
+and [Pinecone guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pinecone.md),
+and [Milvus guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/milvus.md).
 
 ## CLI
 
@@ -209,7 +231,7 @@ cover the remaining commands and endpoints.
 
 ## Status
 
-EmbedFlow v0.2.0 is an alpha release for research and early real-world
+EmbedFlow v0.4.0 is an alpha release for research and early real-world
 testing. T2-v1 is an empirical finite-tail diagnostic, partial rankings can
 differ from fully warm target reranking, and ANN fidelity needs a reference
 comparison to audit.

{embedflow-0.2.0 → embedflow-0.4.0}/README.md

@@ -7,7 +7,7 @@
 EmbedFlow lets a new embedding model serve over candidates from an existing
 vector index while target document vectors are materialized progressively. It
 supports migration analysis, persistent caching, background work, and serving
-through FAISS, Qdrant, pgvector, a CLI, and FastAPI.
+through FAISS, Qdrant, pgvector, Pinecone, Milvus, a CLI, and FastAPI.
 
 [Quickstart](#try-it) · [Documentation](#documentation) · [Research](#research)
 
@@ -52,6 +52,18 @@ For FAISS and the dashboard, add the optional integrations:
 python -m pip install "embedflow[faiss,dashboard]"
 ```
 
+For an existing Pinecone dense index:
+
+```bash
+python -m pip install "embedflow[pinecone]"
+```
+
+For an existing Milvus collection:
+
+```bash
+python -m pip install "embedflow[milvus]"
+```
+
 Qdrant and model-runtime extras are documented in
 [`docs/installation.md`](https://github.com/arnsri33/embedflow/blob/main/docs/installation.md).
 For model-backed analysis, install `embedflow[faiss,models,dashboard]`.
@@ -190,12 +202,16 @@ is measured separately and is `UNKNOWN` until an exact reference is supplied.
 | FAISS | Supported |
 | Qdrant | Supported |
 | pgvector | Supported |
+| Pinecone | Supported |
+| Milvus | Supported |
 
 Backend-specific setup and examples:
 
 - [FAISS](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/faiss.md)
 - [Qdrant](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/qdrant.md)
 - [pgvector](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pgvector.md)
+- [Pinecone](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pinecone.md)
+- [Milvus](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/milvus.md)
 - [Adding a backend](https://github.com/arnsri33/embedflow/blob/main/CONTRIBUTING.md)
 
 ## CLI
@@ -230,7 +246,7 @@ OpenAPI documentation; see
 
 ## Status
 
-EmbedFlow v0.2.0 is an alpha release for research and early real-world
+EmbedFlow v0.4.0 is an alpha release for research and early real-world
 testing.
 
 - T2-v1 reports an empirical finite-tail diagnostic.

{embedflow-0.2.0 → embedflow-0.4.0}/README_PYPI.md

@@ -5,7 +5,7 @@
 EmbedFlow lets a new embedding model serve over candidates from an existing
 vector index while target document vectors are materialized progressively. It
 supports migration analysis, persistent caching, background work, FAISS,
-Qdrant, pgvector, a CLI, and FastAPI.
+Qdrant, pgvector, Pinecone, Milvus, a CLI, and FastAPI.
 
 The full project README and architecture diagram are on
 <https://github.com/arnsri33/embedflow>.
@@ -28,6 +28,18 @@ For an existing PostgreSQL/pgvector table:
 python -m pip install "embedflow[pgvector]"
 ```
 
+For an existing Pinecone dense index:
+
+```bash
+python -m pip install "embedflow[pinecone]"
+```
+
+For an existing Milvus collection:
+
+```bash
+python -m pip install "embedflow[milvus]"
+```
+
 Qdrant and model-runtime extras are documented in the
 [installation guide](https://github.com/arnsri33/embedflow/blob/main/docs/installation.md).
 For model-backed analysis, install `embedflow[faiss,models,dashboard]`.
@@ -129,10 +141,14 @@ for definitions and reproduction details.
 | FAISS | Supported |
 | Qdrant | Supported |
 | pgvector | Supported |
+| Pinecone | Supported |
+| Milvus | Supported |
 
 See the [FAISS guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/faiss.md),
 [Qdrant guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/qdrant.md),
-and [pgvector guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pgvector.md).
+and [pgvector guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pgvector.md),
+and [Pinecone guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/pinecone.md),
+and [Milvus guide](https://github.com/arnsri33/embedflow/blob/main/docs/integrations/milvus.md).
 
 ## CLI
 
@@ -152,7 +168,7 @@ cover the remaining commands and endpoints.
 
 ## Status
 
-EmbedFlow v0.2.0 is an alpha release for research and early real-world
+EmbedFlow v0.4.0 is an alpha release for research and early real-world
 testing. T2-v1 is an empirical finite-tail diagnostic, partial rankings can
 differ from fully warm target reranking, and ANN fidelity needs a reference
 comparison to audit.

{embedflow-0.2.0 → embedflow-0.4.0}/docs/api.md

@@ -49,7 +49,9 @@ request. A partial response scores the available target vectors; it can differ
 from the fully warm ranking.
 
 `/status` includes safe backend metadata. For pgvector this names the
-schema/table and vector contract without returning the DSN or credentials;
+schema/table and vector contract; for Pinecone it names the host/index,
+namespace, dimension, metric, and safe vector counts. Neither backend returns
+the DSN, API key, or credentials;
 `/health` remains a compact liveness response for probes and load balancers.
 
 ## Errors

{embedflow-0.2.0 → embedflow-0.4.0}/docs/configuration.md

@@ -14,7 +14,7 @@ target:
   device: cuda
 
 index:
-  backend: faiss                 # faiss, qdrant, or pgvector
+  backend: faiss                 # faiss, qdrant, pgvector, pinecone, or milvus
   path: ./legacy.index
   ids: ./legacy.index.ids.json   # FAISS sidecar
   metric: cosine
@@ -22,6 +22,10 @@ index:
   # Qdrant fields: url, collection, vector_name, api_key_env
   # pgvector fields: dsn_env, schema, table, id_column, vector_column, text_column
   #                 hnsw_ef_search, ivfflat_probes
+  # Pinecone fields: host (preferred) or index_name, api_key_env, namespace,
+  #                  text_metadata_field
+  # Milvus fields: uri, token_env, database, collection, id_field, vector_field,
+  #                text_field, partition_names, search_params, auto_load
 
 documents:
   path: ./documents.jsonl
@@ -88,6 +92,20 @@ EMBEDFLOW_PGVECTOR_VECTOR_COLUMN
 EMBEDFLOW_PGVECTOR_TEXT_COLUMN
 EMBEDFLOW_PGVECTOR_HNSW_EF_SEARCH
 EMBEDFLOW_PGVECTOR_IVFFLAT_PROBES
+EMBEDFLOW_PINECONE_HOST
+EMBEDFLOW_PINECONE_INDEX_NAME
+EMBEDFLOW_PINECONE_NAMESPACE
+EMBEDFLOW_PINECONE_TEXT_METADATA_FIELD
+EMBEDFLOW_PINECONE_API_KEY_ENV
+EMBEDFLOW_MILVUS_URI
+EMBEDFLOW_MILVUS_TOKEN_ENV
+EMBEDFLOW_MILVUS_DATABASE
+EMBEDFLOW_MILVUS_COLLECTION
+EMBEDFLOW_MILVUS_ID_FIELD
+EMBEDFLOW_MILVUS_VECTOR_FIELD
+EMBEDFLOW_MILVUS_TEXT_FIELD
+EMBEDFLOW_MILVUS_PARTITIONS
+EMBEDFLOW_MILVUS_AUTO_LOAD
 EMBEDFLOW_INDEX_NPROBE
 EMBEDFLOW_DOCUMENTS_PATH
 EMBEDFLOW_CACHE_PATH

{embedflow-0.2.0 → embedflow-0.4.0}/docs/installation.md

@@ -31,6 +31,8 @@ python -m pip install -e .
 | `faiss` | FAISS source-index adapter |
 | `qdrant` | Qdrant client and adapter |
 | `pgvector` | Psycopg 3 binary driver and pgvector adapter |
+| `pinecone` | Official Pinecone Python SDK and adapter |
+| `milvus` | Official pymilvus SDK and adapter |
 | `models` | PyTorch, Transformers, Sentence Transformers, and Hub client |
 | `dashboard` | FastAPI, Uvicorn, and Pydantic |
 | `dev` | Pytest, Ruff, and build tooling |
@@ -59,6 +61,18 @@ Install the optional adapter with `python -m pip install "embedflow[pgvector]"`.
 The adapter connects to an existing table and reads the DSN from the
 environment; see [`integrations/pgvector.md`](integrations/pgvector.md).
 
+## Pinecone
+
+Install the optional adapter with `python -m pip install "embedflow[pinecone]"`.
+Set `PINECONE_API_KEY` in the environment and configure an existing dense
+index host; see [`integrations/pinecone.md`](integrations/pinecone.md).
+
+## Milvus
+
+Install the optional adapter with `python -m pip install "embedflow[milvus]"`.
+Configure an existing collection URI, database, and vector field; see
+[`integrations/milvus.md`](integrations/milvus.md).
+
 ## CPU and GPU
 
 The deterministic demo runs on CPU. Real model serving accepts `--device cpu`

embedflow-0.4.0/docs/integrations/milvus.md

@@ -0,0 +1,165 @@
+# Milvus
+
+EmbedFlow can use an existing Milvus collection as a read-only source
+candidate index. A source query is executed by Milvus, the returned documents
+are reranked by the target model, and target vectors are stored in EmbedFlow's
+local cache. The adapter does not copy or rebuild the source collection.
+
+## Install and connect
+
+Install the optional official SDK (the adapter is tested with pymilvus 2.5.5
+through 3.0.1):
+
+```bash
+python -m pip install "embedflow[milvus]"
+```
+
+For a local standalone server, the URI is normally
+`http://127.0.0.1:19530`. Cloud/Milvus-compatible endpoints can use the same
+`MilvusClient` URI and token semantics. Keep credentials out of YAML:
+
+```bash
+export EMBEDFLOW_MILVUS_TOKEN='user:password-or-cloud-token'
+```
+
+An unauthenticated local server may omit the variable. EmbedFlow never puts a
+token in status, telemetry, generated configuration, or error messages.
+
+Example configuration for a collection that stores its own text:
+
+```yaml
+source:
+  model: sentence-transformers/all-MiniLM-L6-v2
+  dimension: 384
+  device: cuda
+target:
+  model: Qwen/Qwen3-Embedding-0.6B
+  device: cuda
+index:
+  backend: milvus
+  uri: http://127.0.0.1:19530
+  token_env: EMBEDFLOW_MILVUS_TOKEN
+  database: default
+  collection: documents
+  id_field: id
+  vector_field: embedding
+  text_field: content
+  metric: cosine
+  partition_names: []
+  auto_load: false
+documents:
+  path: ./documents.jsonl
+  id_field: id
+  text_field: text
+```
+
+`uri`, `database`, `collection`, `id_field`, `vector_field`, and
+`text_field` are quoted as ordinary SDK arguments; they are not interpolated
+into SQL. `vector_field` should identify a dense `FLOAT_VECTOR` field. The
+first release does not claim support for `SPARSE_FLOAT_VECTOR`, `BINARY_VECTOR`,
+`FLOAT16_VECTOR`, or `BFLOAT16_VECTOR`. If a collection has multiple dense
+vector fields, configure `vector_field` explicitly.
+
+## Text and IDs
+
+There are two supported text layouts:
+
+* Set `text_field` when the Milvus collection has `id`, `embedding`, and a
+  string text field. Search requests ask Milvus for only that field.
+* Keep `text_field` unset and provide the normal EmbedFlow JSONL document
+  store. Candidate retrieval then asks Milvus for IDs only and resolves text
+  through the shared document-store contract.
+
+Milvus `INT64` and `VARCHAR` primary keys are supported. IDs are exposed to
+EmbedFlow as strings, so an INT64 value `42` is represented as `"42"` at the
+application boundary while a VARCHAR value `"42"` remains a string. No numeric
+coercion is performed for VARCHAR IDs.
+
+## Metrics and indexes
+
+Canonical metric names are `cosine`, `inner_product`/`dot`, and
+`l2`/`euclidean`; they map to Milvus `COSINE`, `IP`, and `L2`. Milvus returns
+COSINE/IP similarities (larger is better) and L2 distances (smaller is
+better). EmbedFlow negates L2 distances only, preserving the shared
+higher-is-better `SearchHit.score` convention.
+
+The adapter does not create indexes. Existing HNSW and IVF_FLAT indexes are
+queried when present. Optional native search parameters can be supplied as:
+
+```yaml
+index:
+  search_params:
+    ef: 64       # HNSW; raised to at least top-k when required by Milvus
+    # nprobe: 16 # IVF_FLAT
+```
+
+Native Milvus form (`metric_type` plus `params`) is also accepted. Unknown
+keys are rejected. `ef` is never sent below the requested top-k because HNSW
+servers reject that request.
+
+## Loading, databases, and partitions
+
+EmbedFlow checks collection load state. `auto_load: false` (the conservative
+default) returns an actionable error if the collection is not loaded. Set
+`auto_load: true` only when explicitly permitting EmbedFlow to load this
+collection; EmbedFlow never releases or unloads a collection.
+
+`database` selects the Milvus database. `partition_names` restricts every
+search and text lookup to the listed partitions. A missing partition fails at
+connection time when the server exposes partition metadata. No arbitrary
+Milvus filter expression is exposed because the shared `VectorIndex` contract
+has no portable filter field.
+
+## Commands and audit
+
+```bash
+embedflow doctor --config embedflow.yaml
+embedflow audit-index --config embedflow.yaml
+embedflow analyze --config embedflow.yaml --queries ./probe_queries.jsonl
+embedflow serve --config embedflow.yaml --device cuda
+embedflow search --config embedflow.yaml "your query"
+embedflow status --config embedflow.yaml
+```
+
+`audit-index` reports URI (redacted), database, collection, fields, dense
+vector type, dimension, metric, index type, load state, partitions, row
+count, and a small retrieval/text probe. It intentionally keeps ANN fidelity,
+T2-v1 compatibility, and candidate quality as separate `UNKNOWN`/empirical
+questions; a reachable collection is not proof of migration compatibility.
+
+Normal `doctor`, `audit-index`, `analyze`, `serve`, `search`, `status`, and
+`prewarm` operations are read-only against the configured collection. Fixture
+creation and inserts are confined to `scripts/milvus_fixture.py`, tests, and
+the example setup.
+
+## Local example
+
+The repository includes a small standalone deployment and deterministic
+fixture:
+
+```bash
+cd examples/milvus
+docker compose up -d
+python -m pip install "embedflow[milvus,dashboard]"
+./run_demo.sh
+```
+
+The fixture utility is deliberately explicit and refuses to overwrite an
+existing collection. It can create HNSW, IVF_FLAT, or FLAT test indexes. No
+model weights or credentials are committed.
+
+## Troubleshooting and limitations
+
+* `Milvus support requires ...`: install the `[milvus]` extra in the active
+  environment.
+* `collection ... was not found`: check `uri`, `database`, and collection name.
+* `collection ... is not loaded`: load it administratively or opt in to
+  `auto_load: true`.
+* Dimension/metric errors mean the configured source model contract does not
+  match the existing vector field/index; EmbedFlow does not rewrite it.
+* The adapter uses one synchronous `MilvusClient`; concurrent calls are
+  serialized by a lock. A pool is intentionally not introduced for this
+  release.
+* Milvus-compatible URI/token semantics should work with Zilliz Cloud where
+  supported by the SDK, but Zilliz Cloud has not been independently validated
+  for this release.

embedflow-0.4.0/docs/integrations/pinecone.md

@@ -0,0 +1,157 @@
+# Pinecone
+
+EmbedFlow can use an existing Pinecone dense index as its source candidate
+layer. It queries that index, reranks the returned documents with the target
+embedding model, and materializes target vectors into EmbedFlow's local cache.
+The source index is read-only during normal operation.
+
+## Install
+
+```bash
+python -m pip install "embedflow[pinecone]"
+```
+
+The extra installs the official `pinecone` Python SDK. The base package does
+not import or require it.
+
+## Configure an existing index
+
+Set the API key in the environment, never in YAML:
+
+```bash
+export PINECONE_API_KEY='...'
+```
+
+Use the data-plane host copied from the Pinecone console:
+
+```yaml
+source:
+  model: sentence-transformers/all-MiniLM-L6-v2
+  dimension: 384
+  device: cuda
+
+target:
+  model: Qwen/Qwen3-Embedding-0.6B
+  device: cuda
+
+index:
+  backend: pinecone
+  api_key_env: PINECONE_API_KEY
+  host: my-index-xxxxx.svc.aped-xxxx.pinecone.io
+  namespace: production
+  metric: cosine
+  text_metadata_field: text
+
+documents:
+  # Omit this file when text_metadata_field is present in Pinecone metadata.
+  # If present, it is used as the external text resolver instead.
+  path: ./documents.jsonl
+  id_field: id
+  text_field: text
+```
+
+`host` takes precedence when both `host` and `index_name` are supplied. If a
+host is unavailable, `index_name` can resolve through the control plane; host
+targeting is recommended for production data-plane traffic. The configured
+namespace is passed on every query and fetch. An empty namespace selects
+Pinecone's default namespace.
+
+The source index must be a dense vector index whose dimension and metric match
+the source model contract. Pinecone IDs remain strings, including numeric-
+looking IDs such as `"42"`, UUID-looking values, Unicode, and punctuation.
+
+## Document text
+
+EmbedFlow needs candidate text for target-model encoding. There are two modes:
+
+* Set `text_metadata_field` to read text from Pinecone metadata. Candidate
+  queries request metadata but never request vector values. Missing or
+  non-string text is reported with the candidate ID and field name.
+* Provide the normal JSONL `documents` store when Pinecone contains IDs and
+  vectors only. The external store remains the source of text and avoids
+  duplicating a corpus in Pinecone.
+
+## Commands
+
+```bash
+embedflow doctor --config embedflow.yaml
+embedflow audit-index --config embedflow.yaml
+embedflow analyze --config embedflow.yaml
+embedflow serve --config embedflow.yaml
+embedflow search --config embedflow.yaml "what causes auroras?"
+embedflow status --config embedflow.yaml
+```
+
+`audit-index` separates backend health from retrieval fidelity. It checks SDK
+and credentials, reachability, namespace stats, dimension, metric, dense
+vector type, a small candidate query, and text resolution where possible. A
+successful API call is not an ANN recall or T2-v1 compatibility guarantee.
+
+## Read-only behavior
+
+`doctor`, `audit-index`, `analyze`, `serve`, `search`, `status`, and `prewarm`
+only query the configured index. EmbedFlow does not create or delete indexes,
+upsert or delete vectors, alter metadata, or change namespaces. Test fixtures
+may create temporary indexes explicitly, but production initialization never
+does.
+
+## Metrics and limits
+
+`cosine`, `dot`/`inner_product` (Pinecone `dotproduct`), and
+`l2`/`euclidean` are accepted. Cosine and dot-product scores already rank
+higher-is-better. Pinecone's Euclidean score is squared distance, so EmbedFlow
+negates it to preserve the shared higher-is-better convention. Query `top_k`
+is validated in the range 1–10,000, matching the Pinecone API limit.
+
+## Integrated embedding indexes
+
+Pinecone indexes created with hosted/integrated embedding manage the source
+embedding contract outside EmbedFlow. Unless the source model, dimension, and
+metric can be verified, EmbedFlow reports that contract as unknown; it does not
+assume Pinecone's hosted model is the configured source encoder. The first
+adapter release is intended for standard dense-vector indexes.
+
+## Eventual consistency and tests
+
+Pinecone is eventually consistent after writes. Any test setup that creates a
+temporary fixture must poll `describe_index_stats` with a bounded timeout before
+querying it. EmbedFlow's production adapter performs no writes and therefore
+does not need a write-read delay.
+
+## Troubleshooting
+
+* `Pinecone support requires ...`: install `embedflow[pinecone]` in the active
+  environment.
+* `Environment variable PINECONE_API_KEY is not set.`: export the variable
+  named by `api_key_env`.
+* `Unable to reach configured Pinecone index`: check the host, project, and
+  network policy. Credentials are not printed in this error.
+* Dimension or metric mismatch: compare the source model contract with the
+  existing index configuration; EmbedFlow will not rewrite the index.
+* Missing metadata text: set the correct `text_metadata_field` or provide an
+  external JSONL document store.
+
+## Local smoke fixture
+
+The repository includes a non-destructive smoke helper for an existing index:
+
+```bash
+export PINECONE_API_KEY='...'
+python scripts/pinecone_smoke.py \
+  --host "$PINECONE_INDEX_HOST" \
+  --namespace production
+```
+
+It only describes stats and, when given a query vector, performs a query. It
+never upserts or deletes records. A remote create/query/delete integration is
+opt-in and requires both `EMBEDFLOW_PINECONE_INTEGRATION_TEST=1` and
+`EMBEDFLOW_PINECONE_CREATE_TEST_INDEX=1`; it uses a unique disposable index and
+deletes it in teardown.
+
+## Limitations
+
+The adapter uses one synchronous SDK client. The SDK call itself is safe for
+the modest concurrent serving loads tested by EmbedFlow, but no connection pool
+is introduced. Metadata filters are not exposed because the shared
+`VectorIndex` contract has no portable filter field. Pinecone index provisioning
+and namespace administration remain outside EmbedFlow.

{embedflow-0.2.0 → embedflow-0.4.0}/docs/limitations.md

@@ -1,6 +1,6 @@
 # Limitations and release scope
 
-EmbedFlow v0.2.0 is an alpha release for research and early real-world
+EmbedFlow v0.4.0 is an alpha release for research and early real-world
 testing. The serving path is designed to make migration experiments concrete;
 production rollout still requires application-specific validation.