duxxdb 0.4.0

DuxxDB — the database built for AI agents. Hybrid (vector + BM25 + structured) retrieval, agent-native primitives (MEMORY / TOOL_CACHE / SESSION), embedded-or-server, RESP / gRPC / MCP.

duxxdb — Python bindings

Native Python bindings for DuxxDB, the agent-native hybrid database. Built with PyO3 + maturin against the stable Python ABI (abi3-py38) so a single wheel supports Python 3.8 through 3.13.

Install

From a wheel (recommended)

pip install duxxdb-0.1.0-cp38-abi3-<platform>.whl

(For now, build the wheel locally — see the next section. Public PyPI release lands with v0.1.0.)

Build from source

# Prerequisites: Rust toolchain, Python ≥ 3.8.
pip install --user maturin
cd bindings/python
maturin build --release
pip install --force-reinstall ../../target/wheels/duxxdb-0.1.0-cp38-abi3-*.whl

On Windows + Git Bash the workspace uses the GNU toolchain — see ../../docs/SETUP.md for the WinLibs MinGW prerequisite.

Quickstart

import duxxdb

store = duxxdb.MemoryStore(dim=4)

def embed(text):
    """Replace this with a real embedder (OpenAI / Cohere / local BGE)."""
    import hashlib
    h = int(hashlib.sha1(text.lower().encode()).hexdigest()[:16], 16)
    v = [(h >> (i*4)) & 0xff for i in range(4)]
    norm = sum(x*x for x in v) ** 0.5 or 1.0
    return [x / norm for x in v]

store.remember(key="alice", text="I lost my wallet at the cafe", embedding=embed("wallet"))
store.remember(key="alice", text="My favorite color is blue",     embedding=embed("blue"))

hits = store.recall(key="alice", query="wallet",
                    embedding=embed("wallet"), k=3)
for hit in hits:
    print(f"{hit.score:.4f}  {hit.text}")

Output:

0.0328  I lost my wallet at the cafe
0.0161  My favorite color is blue

API surface

Embedded (native, no server required)

Class	Constructor	Methods
MemoryStore	dim, capacity=100_000	remember(key, text, embedding) -> id, recall(key, query, embedding, k=10) -> [MemoryHit], len(), dim
MemoryHit	(returned by recall)	id, key, text, score
ToolCache	threshold=0.95	put(tool, args_hash, args_embedding, result, ttl_secs=3600), get(tool, args_hash, args_embedding) -> ToolCacheHit | None, purge_expired(), len()
ToolCacheHit	(returned by get)	kind ("exact" or "semantic_near_hit"), similarity, result
SessionStore	ttl_secs=1800	put(session_id, data), get(session_id) -> bytes | None, delete(session_id) -> bool, purge_expired(), len()
PromptRegistry	dim=32, storage=None	put(name, content, metadata=None) -> version, get(name, version_or_tag=None) -> Prompt | None, list(name) -> [Prompt], names() -> [str], tag(name, version, tag), untag(name, tag) -> bool, delete(name, version) -> bool, search(query, k=10) -> [(name, version, score, content)], diff(name, v_a, v_b) -> str
Prompt	(returned by get / list)	name, version, content, tags, metadata (decoded JSON), created_at_unix_ns
PromptHit	(returned by search)	prompt, score

Phase 7 (Phase-7 agent ops, native bindings new in v0.2.1)

Class	Constructor	Methods
CostLedger	dim=32, storage=None	record(tenant, model, tokens_in, tokens_out, cost_usd, ...), query(...), total(tenant), aggregate(group_by, ...), set_budget(tenant, period, amount_usd, warn_pct=0.8, ...), get_budget(tenant), delete_budget(tenant), status(tenant)
CostEntry / Budget	(returned by ledger methods)	typed fields incl. metadata (decoded JSON)
DatasetRegistry	dim=32, storage=None	create(name, schema=None), add(name, rows, metadata=None) -> version, get(name, version_or_tag=None), list(name), names(), tag/untag/delete, sample(name, version, n, split=None), size, splits, search(query, k=10, name_filter=None)
Dataset / DatasetRow	(returned by registry)	typed fields incl. metadata, data, annotations, schema (all decoded JSON)
EvalRegistry	dim=32, storage=None	start(...), score(run_id, row_id, score, output_text="", notes=None), complete(run_id) -> EvalSummary, fail(run_id, reason), get(run_id), scores(run_id), list(dataset_name=None, dataset_version=None), compare(run_a, run_b) -> tuple, cluster_failures(run_id, ...) -> list[tuple]
EvalRun / EvalScore / EvalSummary	(returned by registry)	typed fields incl. metadata, notes (decoded JSON)
ReplayRegistry	storage=None	capture(trace_id, kind, input, ...), get_session(trace_id), list_sessions(), start(source_trace_id, mode="live", ...), step(run_id), record_output(run_id, idx, output), complete/fail, set_replay_trace_id, get_run, list_runs(source_trace_id=None)
ReplaySession / ReplayInvocation / ReplayRun	(returned by registry)	typed fields incl. input, output, metadata (decoded JSON)
TraceStore	storage=None	record_span(trace_id, span_id, name, ...), close_span(span_id, end_unix_ns, status="ok"), get_trace(trace_id), subtree(span_id), thread(thread_id)
Span	(returned by store)	typed fields incl. attributes (decoded JSON)

Module-level: duxxdb.__version__.

Server (typed Python facade over RESP)

When you're running duxx-server as a daemon and want a typed Python surface over every Phase 7 agent-ops primitive (traces, prompts, datasets, evals, replay, cost ledger), install the server extra:

pip install 'duxxdb[server]'

from duxxdb.server import ServerClient

client = ServerClient(url="redis://:<token>@localhost:6379")

# Phase 7.2 — prompt registry
v1 = client.prompts.put("classifier", "you are a refund agent")
prompt = client.prompts.get("classifier", v1)

# Phase 7.3 — dataset registry
client.datasets.create("refunds")
ds_v = client.datasets.add("refunds", [
    {"id": "r1", "text": "I want a refund", "split": "train"},
    {"id": "r2", "text": "Where is my package?", "split": "test"},
])

# Phase 7.4 — evals with summary stats + failure clustering
run_id = client.evals.start(
    dataset_name="refunds",
    dataset_version=ds_v,
    model="gpt-4o-mini",
    scorer="llm_judge",
    prompt_name="classifier",
    prompt_version=v1,
)
client.evals.score(run_id, row_id="r1", score=0.9, output_text="REFUND")
client.evals.score(run_id, row_id="r2", score=0.1, output_text="REFUND")
summary = client.evals.complete(run_id)
print(summary.mean, summary.pass_rate_50)

# Phase 7.6 — cost ledger
client.cost.record(tenant="acme", model="gpt-4o-mini",
                   tokens_in=120, tokens_out=80, cost_usd=0.0023)
print(client.cost.total("acme"))

Namespace	Wraps	Methods (highlights)
client.trace	TRACE.* (6 cmds)	record, close, get, subtree, thread, search
client.prompts	PROMPT.* (9 cmds)	put, get, list, names, tag, untag, delete, search, diff
client.datasets	DATASET.* (13 cmds)	create, add, get, sample, size, splits, search, from_recall, …
client.evals	EVAL.* (9 cmds)	start, score, complete, get, scores, list, compare, cluster_failures
client.replay	REPLAY.* (12 cmds)	capture, start, step, record, complete, diff, list_runs, …
client.cost	COST.* (10 cmds)	record, query, aggregate, total, set_budget, status, alerts, cluster_expensive

All return types are plain dataclasses decoded from the server's JSON responses. The raw redis.Redis client is exposed as client.raw for anything not yet wrapped.

PromptRegistry (embedded, native bindings): versioned prompts with semantic search

New in v0.2.0: native PyO3 bindings for the prompt registry, so you can use it embedded — no server, no redis-py — straight from a notebook or batch script. The same Rust crate that powers duxx-server's PROMPT.* commands is exposed directly.

import duxxdb

# In-memory (default; matches v0.1.x in-memory feel).
r = duxxdb.PromptRegistry(dim=16)
v1 = r.put("classifier", "You are a refund agent.", metadata={"author": "alice"})
v2 = r.put("classifier", "You are a friendly refund agent.")
r.tag("classifier", v2, "prod")

p = r.get("classifier", "prod")     # resolves the tag
print(p.version, p.content, p.metadata)

# Durable (rows + tags + monotonic counter survive process exit).
r = duxxdb.PromptRegistry(dim=16, storage="redb:./prompts.redb")
r.put("greeting", "Hello! How can I help today?")
# ... process dies ...
r = duxxdb.PromptRegistry(dim=16, storage="redb:./prompts.redb")
assert r.get("greeting").content == "Hello! How can I help today?"

# Semantic search across the catalog.
for hit in r.search("hello", k=3):
    print(hit.score, hit.prompt.name, hit.prompt.content)

The HNSW vector index is rebuilt on open by re-embedding every persisted prompt — fine for the typical prompt-catalog scale (<1000 rows). Larger Phase 7 primitives (datasets, evals) keep shipping through duxxdb.server.ServerClient for now; native PyO3 bindings for them land progressively through v0.2.x.

Phase 7 (embedded, native) — v0.2.1 worked example

The five Phase 7 primitives that shipped through the RESP facade in v0.1.3 are now available natively too — no duxx-server, no redis-py. Same Rust crates the server uses.

import duxxdb

# Cost ledger
cost = duxxdb.CostLedger(dim=16, storage="redb:./cost.redb")
cost.record(tenant="acme", model="gpt-4o-mini",
            tokens_in=120, tokens_out=80, cost_usd=0.0023,
            input_text="please refund order #9910")
cost.set_budget("acme", "monthly", 100.0, warn_pct=0.8)
print(cost.total("acme"), cost.status("acme"))

# Dataset registry
ds = duxxdb.DatasetRegistry(dim=16, storage="redb:./ds.redb")
ds.create("refunds", schema={"columns": ["text", "label"]})
v = ds.add("refunds", [
    {"id": "r1", "text": "I want a refund", "split": "train"},
    {"id": "r2", "text": "Where is my package?", "split": "test"},
])

# Eval registry
evals = duxxdb.EvalRegistry(dim=16, storage="redb:./evals.redb")
rid = evals.start(dataset_name="refunds", dataset_version=v,
                  model="gpt-4o-mini", scorer="llm_judge")
evals.score(rid, row_id="r1", score=0.9, output_text="REFUND")
evals.score(rid, row_id="r2", score=0.1, output_text="REFUND")
summary = evals.complete(rid)
print(summary.mean, summary.p99, summary.pass_rate_50)

# Replay
replay = duxxdb.ReplayRegistry(storage="redb:./replay.redb")
replay.capture(trace_id="t1", kind="llm_call",
               input={"messages": [{"role": "user", "content": "hi"}]},
               output={"role": "assistant", "content": "hello"})
run_id = replay.start("t1", mode="live")

# Trace store
trace = duxxdb.TraceStore(storage="redb:./trace.redb")
trace.record_span(trace_id="t1", span_id="root", name="agent.turn",
                  attributes={"user": "alice"}, status="ok",
                  start_unix_ns=1_000_000_000)
spans = trace.get_trace("t1")
print(spans[0].name, spans[0].attributes)

Every JSON-shaped field (metadata, attributes, data, annotations, notes, input, output, schema) round-trips transparently to native Python dicts/lists.

ToolCache: semantic-near-hit demo

cache = duxxdb.ToolCache(threshold=0.95)

# Cache the result of an expensive web_search call.
cache.put(tool="web_search", args_hash=hash("what is rust?"),
          args_embedding=embed("what is rust?"),
          result=b"A systems programming language ...",
          ttl_secs=600)

# Later, a paraphrased query — different hash, similar embedding.
hit = cache.get("web_search",
                args_hash=hash("describe rust"),
                args_embedding=embed("describe rust"))

if hit and hit.kind == "semantic_near_hit":
    print(f"cache hit by paraphrase, similarity={hit.similarity:.3f}")
    answer = hit.result.decode() if isinstance(hit.result, bytes) else bytes(hit.result).decode()

What's missing today

• Single-file redb: mode for --phase7-storage. Today only dir:<directory> is supported (one redb file per primitive). Single-file mode requires sharing one Arc<Database> across all six backends — queued for a future v0.2.x.
• Subscriptions (MemoryStore.subscribe()) — Phase 4.5; the Rust / RESP servers already support this, the Python wrapper just needs to bridge tokio::broadcast::Receiver into a Python iterator.
• Async API — currently sync only. async def wrappers are planned.
• Numpy / typed-array fast path — embeddings are currently list[float]. Phase 3.5 will accept np.ndarray[float32] directly via numpy::PyArrayLike.

See ../../docs/ROADMAP.md for the broader plan.

License

Apache 2.0. See ../../LICENSE.