WARM SDK

Python HTTP client and framework adapters for a WARM service.

Website · Documentation · Source · License

The client talks to warm-server, which ships in the WARM repository. The server stores vectors you supply and exposes document and query endpoints. See the server guide to run it.

Installation

uv add warm-sdk

Framework adapters are optional extras. Install one, or all of them:

uv add "warm-sdk[langchain]"
uv add "warm-sdk[all]"

The earlier standalone langchain-holographic distribution is retired. For LangChain, install warm-sdk[langchain] and import WarmVectorStore from writerslogic_warm_sdk.langchain. This adapter connects to warm-server; update the old langchain_holographic import and its local-store configuration when migrating.

The fast extra adds orjson for request encoding, which cuts the per-query cost of sending an embedding:

uv add "warm-sdk[fast]"

Requires Python 3.10 or newer.

Quick Start

from writerslogic_warm_sdk import WarmClient

client = WarmClient(url="http://localhost:8080", api_key="...")

client.add_documents([
    {"id": "doc-1", "text": "Paris is the capital of France.", "vector": [0.12, -0.03, 0.88], "metadata": {"lang": "en"}},
])

response = client.query(query_vector=[0.10, -0.01, 0.90], top_k=3, filter={"lang": "en"})
print(response["matches"])
client.delete_document("doc-1")
client.close()

AsyncWarmClient has the same methods as coroutines; close it with await client.aclose(). Embeddings come from your own model: the SDK sends vectors, it does not compute them.

After bulk ingestion, client.maintain_document_index() explicitly builds the server's exact dense document index and returns {"documentIndexTrained": true} when the store is eligible. Smaller or unsupported stores return false and use exact scanning. Use await client.maintain_document_index() with the async client. Rebuild after updates or restart; the index is a runtime cache. Searches preserve the existing cosine results, and metadata filters continue to apply before ranking. This method requires an warm-server version exposing POST /api/v1/index/maintain.

Source Evidence

complete_evidence packs complete source observations in your retrieval order. Each observation contains a turn handle, session handle, observation time, role, and verbatim text. Supply opaque caller IDs as lowercase SHA-256 hex handles, and keep their association with your original records.

from hashlib import sha256
from writerslogic_warm_sdk import complete_evidence, decode_source_packet

def handle(identity):
    return sha256(identity.encode("utf-8")).hexdigest()

session = handle("session-7")
candidates = [
    [(handle("turn-1"), session, "2026-10-13T10:00:00Z", "user",
      "I have two aquarium tanks.")],
    [(handle("turn-2"), session, "2026-10-13T10:00:00Z", "assistant",
      "You mentioned two aquarium tanks.")],
]

selection = complete_evidence(candidates, selected=[0], budget_bytes=2048)
observations = decode_source_packet(selection.packet)
assert observations == [candidates[index] for index in selection.selected]

The indices in selected are mandatory; packing raises ValueError if they exceed the budget. Additional complete groups fit in candidate order, with shared text and session metadata stored once. Decoding preserves every occurrence, role, observation time, and source text. The budget counts UTF-8 packet bytes; measure reader tokens separately. Packets contain source observations and no inferred answer or event date. These helpers run locally and make no service calls.

Integrations

Vector Masking

Passing zero_trust_key masks every embedding client-side with a keyed signed permutation (VectorMask, scrypt-derived). The server can still rank vectors because inner products and cosine similarity are preserved exactly. This is obfuscation, not encryption: the server learns all pairwise similarities, and known plain/masked pairs reveal the key for the coordinates they cover. Pass a per-collection mask_salt (at least 16 bytes) shared by all clients of that collection. Masks produced by 0.1.0/0.1.1 are not compatible and were not secure.

License

Apache-2.0. See the LICENSE.