OCIEmbeddings hardcodes input_type="SEARCH_QUERY" on embed_query → 1024-dim vectors, breaks RAG against SEARCH_DOCUMENT-indexed (1536-dim) tables when using cohere.embed-v4.0

[fede-kamel]

Summary

locus.rag.embeddings.oci.OCIEmbeddings.embed_query() (around line 370 in 0.2.0b25) hardcodes input_type="SEARCH_QUERY" on the underlying EmbedTextDetails request, ignoring self.oci_config.input_type.

For cohere.embed-v4.0 on OCI Generative AI, the two input_types return different output dimensions:

input_type	output dim
SEARCH_DOCUMENT	1536
SEARCH_QUERY	1024

(This asymmetry is OCI/Cohere v4 behaviour, not Locus's fault — but it interacts badly with the hardcoded override.)

embed_documents() correctly uses SEARCH_DOCUMENT, so an ingestion path that calls embed_documents populates a vector column at 1536 dim. Then a search via embed_query returns a 1024-dim vector. VECTOR_DISTANCE(embedding /*1536*/, :qvec /*1024*/, COSINE) raises:

ORA-12801: error signaled in parallel query server P000, instance 1
ORA-51803: Vector dimension count must match the dimension count
specified in the column definition (expected 1536 dimensions,
specified 1024 dimensions).

Net result: every RAG query against a v4-indexed table errors out, with no warning at config / startup time.

Reproduction

import asyncio
from locus.rag import OCIEmbeddings, OracleVectorStore, RAGRetriever
from locus.rag.embeddings.oci import OCIEmbeddingConfig

async def main():
    e = OCIEmbeddings(
        oci_config=OCIEmbeddingConfig(
            model_id="cohere.embed-v4.0",
            compartment_id="…",
            profile_name="DEFAULT",
            input_type="SEARCH_DOCUMENT",      # explicit; ignored at query time
        )
    )
    r = await e.embed_query("luxury hotels in Greece")
    print(len(r.embedding))   # → 1024  (expected 1536)

asyncio.run(main())

A retriever built on top of this against a VECTOR(1536, FLOAT32) table errors at search time with the ORA-12801 above.

Affected code

• locus/rag/embeddings/oci.py:370 (or thereabouts) — embed_details = EmbedTextDetails(... input_type="SEARCH_QUERY", ...) inside embed_query().

Suggested fix

Two non-exclusive options:

1. Honor self.oci_config.input_type when set: only fall through to "SEARCH_QUERY" when the user hasn't supplied one. Lets callers pin both embed_documents and embed_query to the same input_type when they need matching dims.

   input_type = self.oci_config.input_type if self.oci_config.input_type else "SEARCH_QUERY"

2. Expose output_dimensions on OCIEmbeddingConfig: Cohere v4 in OCI accepts output_dimensions on the embed request; OCI's EmbedTextDetails exposes it. Letting callers pin a target dim (e.g. 1024 or 1536) makes the request symmetric regardless of input_type.

Either fix unblocks v4-indexed RAG. The first is the smaller change.

Environment

• locus-sdk[oci,server] 0.2.0b25
• oci 2.176.0
• cohere.embed-v4.0 via OCI Generative AI, us-chicago-1
• Oracle 26ai (Autonomous Database, free tier), VECTOR(1536, FLOAT32) + HNSW cosine index
• Python 3.13

Related

• OracleBackend silently drops wallet_password="", breaks auto-login wallets (Pydantic v2 SecretStr is falsy) #289 (just-fixed wallet_password truthy guard) — same RAG path, different bug.

[fede-kamel]

Root cause correction (verified against live OCI)

Reproduced on OCI GenAI (us-chicago-1). The input_type hypothesis doesn't hold —
input_type does not change the output dimension:

model	SEARCH_DOCUMENT	SEARCH_QUERY
cohere.embed-v4.0	1536	1536
cohere.embed-english-v3.0	1024	1024
cohere.embed-multilingual-v3.0	1024	1024
cohere.embed-english-light-v3.0	384	384

The actual lever is Cohere v4 Matryoshka output_dimensions, which Locus never set:

v4 default (no output_dimensions): 1536
v4 output_dimensions=1024:         1024   ← the reported 1024
v4 output_dimensions={256,512,1536}: {256,512,1536}

So ORA-51803 arises when the stored column and the query vector were produced at
different output_dimensions (e.g. a foreign ingestion pipeline / older default vs.
Locus's implicit 1536), not from SEARCH_QUERY vs SEARCH_DOCUMENT.

Fix in #293: add OCIEmbeddingConfig.output_dimensions (forwarded only when set, so v3
is untouched) and make config.dimension honor it, so every embed path and the vector
column stay on one dimension. The input_type hardcode in embed_query/embed_documents
(and the dead original_type local) is also fixed — real, but not the dimension cause.