1. Overview

It is responsible for:

• validating configuration before I/O,
• instantiating LLMs, stores, memory, tools, and optional HTTP server,
• returning a runtime pair: (agent, teardown).

All fluent methods return self for chaining. build() is the only async public method.

2. Public API Coverage Checklist

This document covers every public constructor/method/property on AgentBuilder.

2.1 Constructor and Core Fluent Methods

2.2 Escape Hatch Injection Methods

3. Constructor and Fluent Methods

3.1 __init__(name, agent_id, personality)

Creates a builder with required identity fields.

Immediate validation:

• name cannot be empty/whitespace,
• agent_id cannot be empty/whitespace,
• personality cannot be empty/whitespace.

Failure raises ConfigError immediately.

3.2 llm(config: LLMConfig)

Sets the primary LLM used for:

• core reasoning,
• flow selection,
• message enhancement,
• summarization.

If neither .llm(...) nor .with_llm(...) is provided, build() fails validation.

3.3 script_gen_llm(config: LLMConfig)

Optional dedicated LLM for script generation in tool orchestration.

If omitted, script generation falls back to the primary LLM.

3.4 embeddings(...)

Sets automatic embedding model preparation. During build(), AgentBuilder can:

• download tokenizer,
• export ONNX model,
• quantize ONNX (INT8) when requested,
• patch memory config paths automatically.

Parameters:

• model_id default "BAAI/bge-small-en-v1.5"
• output_dir default empty (resolved to ~/.jazzmine/models/<slug>/)
• tokenizer_only default False
• quantized default True
• opset default 17
• force_rebuild default False

Notes:

• tokenizer_only=True skips ONNX export.
• If .embeddings() is called without .memory(), builder logs a warning but does not fail.

3.5 memory(...)

Enables Qdrant-backed long-term memory (episodic/procedural/semantic).

Important behavior:

• If .embeddings() is used, tokenizer_path and (for local mode) model_dir are auto-resolved.
• Without .embeddings(), you must provide the embedding backend manually.

Backend modes:

• automatic: call .embeddings() before .memory().
• manual local: set model_dir.
• manual remote: set embed_api_key (and optionally provider/base URL/model name).

Qdrant automation:

• if qdrant_auto_start=True, qdrant_api_key is None, and URL host is local (localhost, 127.0.0.1, 0.0.0.0, ::1), builder can auto-start a local Docker Qdrant container.

3.6 storage(config: StorageConfig)

Sets message storage backend using one of:

• JsonStorage,
• PostgresStorage,
• MongoDBStorage.

Default is JsonStorage(). If JSON storage path is empty, builder creates a temporary file and registers cleanup in teardown.

3.7 sandbox(...)

Registers one sandbox spec per call.

Use this to define:

• runtime/python version,
• resource limits,
• networking allowlists,
• file mounts,
• package list,
• secrets,
• execution mode and pool sizing.

Call multiple times for multiple named sandboxes.

3.8 flows(flows: list[Any])

Registers flow definitions to be synchronized during build when memory is enabled.

Important: flow registry is weak-reference-backed, so keep a strong reference to your flow list in caller scope.

3.9 domain_terms(terms: list[DomainTerm])

Defines semantic glossary entries.

Each entry must be a 5-tuple: (key, value, category, aliases, description).

Entries are diff-synced to semantic memory during build when memory is active.

3.10 settings(**kwargs)

Overrides AgentSettings fields by name.

Unknown key behavior:

• raises ConfigError with list of valid settings fields.

3.11 version(version: str)

Sets optional version string used by startup display.

3.12 on_intermediate(callback: Callable)

Registers callback for sandbox intermediate events.

3.13 logging(config: dict)

Enables structured runtime logging via AgentLogger.from_config(...).

If configured, logger startup/shutdown is integrated into build/teardown.

3.14 security(config: SecurityConfig)

Enables SecurityGuard with input/output moderation and optional file sanitizer.

Notable rules:

• input_moderator and toxicity_detector are mutually exclusive.
• if .security() is not called, runtime uses NOOP_GUARD.

3.15 server(config: ServerConfig)

Attaches HTTP server lifecycle to the built agent.

Behavior:

• server starts at end of build(),
• server stop callback is registered in teardown,
• effective server instance is available via builder.agent_server.

Default endpoint set (from ServerConfig):

• POST {conversations_endpoint} create conversation
• GET {conversations_endpoint} list conversations for user
• GET {conversations_endpoint}/search search conversations
• GET {conversations_endpoint}/{conversation_id}/messages list messages
• PATCH {conversations_endpoint}/{conversation_id} update metadata
• DELETE {conversations_endpoint}/{conversation_id} delete conversation
• POST {chat_endpoint} full response
• POST {chat_endpoint}/stream SSE response stream
• GET {health_endpoint} health probe
• GET {info_endpoint} agent metadata

3.16 agent_server property

Returns AgentServer | None.

Typical use:

• after build(), inspect builder.agent_server.effective_port when using port=0 in tests.

4. Escape Hatches and Injection Strategy

Escape hatches let you bypass default construction and inject pre-built components.

Use cases:

• unit/integration tests,
• custom implementations,
• shared connection pools/registries,
• deterministic runtime wiring.

Coverage:

• with_llm, with_script_gen_llm
• with_message_store, with_wm_store, redis_wm
• with_registry, with_pool
• with_enhancer, with_summarizer, with_flow_selector
• with_episodic_memory, with_semantic_memory

5. Validation Model (_validate)

Validation executes synchronously at the top of build() before opening runtime resources.

It accumulates all failures and raises one ConfigError containing bulleted details.

5.1 LLM Validation

• requires either .llm(...) or .with_llm(...).
• validates provider config objects through _validate_llm_cfg.

5.2 Embedding and Memory Cross-Validation

Checks include:

• embeddings.model_id non-empty,
• embeddings.opset >= 11,
• output dir is not an existing file,
• tokenizer/backend consistency rules,
• embedding mode conflicts (auto/local/remote).

Memory checks include:

• qdrant_url non-empty,
• vector_size >= 1,
• quantization in {None, 4, 8},
• valid distance_metric in {cosine, euclidean, dot},
• positive shard/replication/index flush limits,
• max_batch >= 2,
• remote provider allowlist validation when API mode is used.

5.3 Storage Validation

• Postgres: non-empty DSN, pool_min >= 1, pool_max >= pool_min.
• MongoDB: non-empty URI and DB name.

5.4 Sandbox Validation

Per sandbox spec:

• non-empty unique name,
• resource lower bounds (memory, pids, timeout, output cap, CPU, pool, scratch),
• swap rule (-1 or > memory_mb),
• port ranges (1..65535),
• non-empty file mount paths,
• no blank pip package entries,
• execution mode in {plan, interactive}.

5.5 Settings Validation

Checks include:

• default_execution_mode in {plan, interactive},
• positive turn/retry limits,
• enhancer_history_window >= 2,
• summarizer_trigger >= 1,
• episode_overlap < max_episode_size,
• pool_max_overflow >= 0,
• positive flow retrieval limits.

5.6 Domain Terms Validation

Each term must be a tuple/list of length 5.

5.7 Security Validation

Checks include:

• input_moderator XOR toxicity_detector,
• required interface methods (classify, predict, sanitize) when objects are provided,
• threshold ranges in [0.0, 1.0],
• positive moderation timeout,
• non-empty block messages.

5.8 Server Validation (_validate_server_cfg)

Checks include:

• port in 1..65535, non-empty host,
• all endpoint paths start with /,
• endpoint uniqueness,
• no collision with auto stream path ${chat_endpoint}/stream,
• positive request limits,
• SSL cert/key pair rules and file existence,
• valid log level,
• valid backlog/keep-alive/concurrency settings,
• CORS credential + wildcard-origin conflict prevention,
• non-empty CORS origin list,
• non-blank API key string if set.

6. Build Lifecycle (await build())

Execution order is deterministic.

1. run _validate().
2. initialize startup display and register teardown banner callbacks.
3. if configured, run embedding setup before memory construction.
4. construct primary/script LLMs (or use injected instances).
5. prepare tool registry and compute registry change report.
6. create sandbox pool if sandboxes/tools are present and no pool injected.
7. initialize message store (with temp JSON file support).
8. initialize working memory store (default in-process unless injected).
9. initialize long-term memory (_build_memory) with optional Qdrant auto-start.
10. sync domain terms and flows when long-term memory is active.
11. construct flow selector, enhancer, summarizer.
12. build AgentConfig and then the runtime agent.
13. optionally start HTTP server and expose builder.agent_server.
14. register teardown callbacks and return (agent, teardown).

7. Runtime/Dependency Details

7.1 Embedding Setup Internals

_build_embeddings():

• checks cache (tokenizer.json and ONNX where applicable),
• checks required Python packages (transformers, and for ONNX export also torch, onnxruntime-tools),
• executes tokenizer/ONNX export in thread executor,
• patches self._mem_cfg.tokenizer_path, and local model_dir/quantized when relevant.

7.2 Memory Runtime Internals

_build_memory():

• validates onnxruntime availability/version (requires >=1.23.0),
• attempts to resolve ORT shared library path from installed wheel,
• imports Rust extension module memory,
• creates QdrantManager, ensures collections, then builds:
• EpisodicMemory,
• ProceduralMemory,
• SemanticMemory.

If .memory() is not configured, builder returns no-op episodic/semantic memory stubs.

8. Error Handling and Failure Modes

8.1 ConfigError

Thrown by constructor (required identity fields) and by _validate() aggregate failures.

8.2 Embedding Setup Failures

Embedding export errors are wrapped into clear ConfigError messages including:

• model,
• output dir,
• common causes (network, disk, missing deps, invalid model id).

8.3 Memory Runtime Import/Version Failures

Common explicit failures:

• missing onnxruntime,
• too-old onnxruntime,
• missing Rust extension module memory.

8.4 HTTP Server Startup Failures

Server startup failures are raised as RuntimeError with host/port context.

9. Practical Examples

9.1 Minimal Agent (No Long-Term Memory)

import asyncio

from jazzmine.core.builder import AgentBuilder, OpenAILLMConfig


async def main() -> None:
    builder = (
        AgentBuilder(
            name="Aria",
            agent_id="aria-local",
            personality="Helpful and concise assistant.",
        )
        .llm(OpenAILLMConfig(model="gpt-4o-mini", api_key="sk-..."))
    )

    agent, teardown = await builder.build()
    try:
        result = await agent.chat(
            user_id="u1",
            conversation_id="c1",
            content="Hello",
        )
        print(result.response)
    finally:
        await teardown()


if __name__ == "__main__":
    asyncio.run(main())

9.2 Local Embeddings + Auto Qdrant

import asyncio

from jazzmine.core.builder import AgentBuilder, OpenAILLMConfig


async def main() -> None:
    builder = (
        AgentBuilder("Aria", "aria-mem", "Memory-enabled assistant")
        .llm(OpenAILLMConfig(model="gpt-4o-mini", api_key="sk-..."))
        .embeddings(
            model_id="BAAI/bge-small-en-v1.5",
            quantized=True,
        )
        .memory(
            qdrant_url="http://localhost:6334",
            qdrant_auto_start=True,
            vector_size=384,
        )
    )

    agent, teardown = await builder.build()
    try:
        await agent.chat(user_id="u1", conversation_id="c1", content="Remember this")
    finally:
        await teardown()


if __name__ == "__main__":
    asyncio.run(main())

9.3 Remote Embeddings (Tokenizer-Only Local Prep)

import asyncio

from jazzmine.core.builder import AgentBuilder, OpenAILLMConfig


async def main() -> None:
    builder = (
        AgentBuilder("Aria", "aria-remote-embed", "Uses remote embedding API")
        .llm(OpenAILLMConfig(model="gpt-4o-mini", api_key="sk-..."))
        .embeddings(tokenizer_only=True)
        .memory(
            qdrant_url="http://localhost:6334",
            embed_api_key="sk-remote-embed",
            embed_provider="openai",
            vector_size=1536,
        )
    )

    agent, teardown = await builder.build()
    try:
        await agent.chat(user_id="u1", conversation_id="c1", content="Store this fact")
    finally:
        await teardown()


if __name__ == "__main__":
    asyncio.run(main())

9.4 Server + Security + CORS

import asyncio

from jazzmine.core.builder import (
    AgentBuilder,
    OpenAILLMConfig,
    SecurityConfig,
    ServerConfig,
    CORSConfig,
)


async def main() -> None:
    builder = (
        AgentBuilder("Aria", "aria-http", "Served assistant")
        .llm(OpenAILLMConfig(model="gpt-4o-mini", api_key="sk-..."))
        .security(SecurityConfig(fail_open=True))
        .server(
            ServerConfig(
                host="127.0.0.1",
                port=8000,
                api_key="super-secret-token",
                cors=CORSConfig(
                    origins=["https://app.example.com"],
                    allow_credentials=True,
                ),
            )
        )
    )

    agent, teardown = await builder.build()
    try:
        # Useful in tests, especially when using ServerConfig(port=0)
        print(builder.agent_server.effective_port)
        await asyncio.sleep(5)
    finally:
        await teardown()


if __name__ == "__main__":
    asyncio.run(main())

9.5 Testing With Escape Hatches

import asyncio

from jazzmine.core.builder import AgentBuilder, JsonStorage
from jazzmine.core.llm.types import LLMResponse, LLMUsage


class FakeLLM:
    model = "fake"

    def generate(self, messages, *, stop=None, **kwargs):
        return LLMResponse(text="ok", usage=LLMUsage(), model=self.model)

    def stream(self, messages, *, stop=None, **kwargs):
        yield "ok"

    async def agenerate(self, messages, *, stop=None, **kwargs):
        return LLMResponse(text="ok", usage=LLMUsage(), model=self.model)

    async def astream(self, messages, *, stop=None, **kwargs):
        yield "ok"

    def close(self):
        return None

    async def aclose(self):
        return None


class FakeStore:
    async def store_message(self, *args, **kwargs):
        return None

    async def flag_message(self, *args, **kwargs):
        return None

    async def close_conversation(self, *args, **kwargs):
        return None

    async def list_conversations(self, *args, **kwargs):
        return []

    async def list_messages(self, *args, **kwargs):
        return []

    async def close(self):
        return None


async def main() -> None:
    fake_llm = FakeLLM()
    fake_store = FakeStore()

    builder = (
        AgentBuilder("TestAgent", "test-agent", "Testing profile")
        .storage(JsonStorage(path="./test_store.json"))
        .with_llm(fake_llm)
        .with_message_store(fake_store)
    )

    agent, teardown = await builder.build()
    try:
        # Run focused assertions on wiring without exercising full chat pipeline.
        assert agent is not None
    finally:
        await teardown()


if __name__ == "__main__":
    asyncio.run(main())

10. Operational Guidance

• Always keep a strong reference to your flow instances/list after .flows(...).
• Prefer .embeddings() + .memory() together to avoid path mismatch errors.
• For remote embeddings, use .embeddings(tokenizer_only=True) plus embed_api_key in memory config.
• Call await teardown() in finally blocks to guarantee cleanup.
• Use escape hatches in tests to avoid external dependencies.