Akasha

Akasha is a Python toolkit for document question answering, retrieval-augmented generation (RAG), native tool-calling agents, summaries, and long-term semantic memory.

It provides one consistent interface for remote and local model workflows while keeping provider-specific integrations behind model aliases such as openai:, gemini:, anthropic:, and ollama:.

Chinese manual: https://tea9297.github.io/akasha/
Current package version: 1.5

What Akasha provides

Capability	Public entry point	Purpose
Chat / QA	`akasha.ask()`	Ask a model a question, optionally with documents or web information
Agents	`akasha.agents()`	Use LangChain-native tool calling, streaming, thinking events, Skills, and MCP tools
RAG	`akasha.RAG()`	Load documents, create embeddings, search Chroma, and generate an answer
Summaries	`akasha.summary()`	Summarize text, files, or URLs with `map_reduce` or `refine`
Long-term memory	`MemoryManager`	Store and retrieve semantic memories with Chroma

Installation

Python 3.11 or 3.12 is recommended.

Lightweight installation

Use light for remote chat models and remote embeddings:

uv venv --python 3.11

# macOS / Linux
source .venv/bin/activate

# Windows PowerShell
# .venv\Scripts\Activate.ps1

uv pip install "akasha-terminal[light]"

light keeps Chroma-backed RAG and memory workflows, but does not include the local HuggingFace / Torch model stack.

Full installation

Use full when you need local embeddings, local HuggingFace models, local Llama/GPTQ models, reranking, or BERTScore:

uv pip install "akasha-terminal[full]"

The practical difference is:

Installation	Chat models	Embeddings	Vector store	Local ML / rerank
`light`	Remote providers	Remote APIs	Local Chroma	No
`full`	Remote and local providers	Remote and local	Local Chroma	Yes

light does not include Torch, Transformers, Sentence-Transformers, or onnxruntime; these local-model dependencies are part of full.

Editable installation for development

uv pip install -e ".[light,dev]"

For the complete local-model stack:

uv pip install -e ".[full,dev]"

Configure a model provider

Set provider credentials in the environment or in a .env file. Never commit .env files or API keys.

OPENAI_API_KEY=your_key
GEMINI_API_KEY=your_key
ANTHROPIC_API_KEY=your_key

# Optional Azure OpenAI-compatible endpoint
AZURE_OPENAI_API_KEY=your_key
AZURE_OPENAI_BASE_URL=https://your-resource.openai.azure.com/

# Optional Ollama endpoint
OLLAMA_API_BASE=http://localhost:11434

Supported chat model aliases include:

openai:gpt-4o
gemini:gemini-2.5-flash
anthropic:claude-3-5-sonnet-latest
ollama:qwen3:8b
azure:your-deployment-name

Ollama can also target another host:

ollama:http://192.168.1.10:11434@qwen3:8b

The same public interfaces accept an already configured LangChain ChatModel when provider-specific configuration is needed.

Quick start: chat

import akasha

qa = akasha.ask(model="gemini:gemini-2.5-flash")
answer = qa("What is retrieval-augmented generation?")
print(answer)

ask(stream=False) returns a final str.

Quick start: RAG

RAG uses a local Chroma store and an embedding model selected independently from the chat model:

import akasha

rag = akasha.RAG(
    model="gemini:gemini-2.5-flash",
    embeddings="gemini:gemini-embedding-001",
)

answer = rag("./docs", "What are the main ideas in these documents?")
print(answer)

Typical embedding aliases include:

openai:text-embedding-3-small
gemini:gemini-embedding-001
hf:BAAI/bge-base-en-v1.5       # full installation

In light, use remote embeddings. Local HuggingFace / Sentence-Transformers embeddings require full.

Quick start: agents and tools

Agents use LangChain 1.3+ native tool calling. A custom Python function can be exposed as a tool with create_tool():

import akasha


def today_f() -> str:
    return "The tool was called successfully."


today_tool = akasha.create_tool(
    "Return the current date or a short status message.",
    today_f,
    "today_status",
)

agent = akasha.agents(
    model="gemini:gemini-2.5-flash",
    tools=[today_tool],
)

print(agent("Use the available tool and report its result."))

Create the agent once and reuse it for multiple questions. Rebuilding an agent for every question repeats provider initialization costs.

Streaming events

Non-streaming calls return a string. Streaming agents return JSON-serializable event dictionaries:

agent = akasha.agents(
    model="gemini:gemini-2.5-flash",
    tools=[],
    stream=True,
    thinking=True,
)

for event in agent("Explain the difference between a vector store and an embedding model."):
    if event["type"] == "thinking":
        print("[thinking]", event["data"])
    elif event["type"] == "tool":
        print("[tool]", event["data"])
    elif event["type"] == "answer":
        print(event["data"], end="", flush=True)

The event types are:

Event	Meaning
`answer`	A chunk of the final answer
`thinking`	Provider reasoning/thinking content, when available and enabled
`tool`	A tool or Skill result

ask(stream=True, thinking=False) currently yields text chunks. ask(stream=True, thinking=True) yields answer and optional thinking events.

Skills and MCP

Agents can load Skills from a Skill directory containing SKILL.md:

agent = akasha.agents(
    model="gemini:gemini-2.5-flash",
    skills=["examples/examples_skills/python-repl-skill"],
)

Skills can provide instructions, resources, and allowlisted tool bundles. Skill tools are surfaced through normal tool events.

MCP tools can be discovered with langchain-mcp-adapters, normalized with akasha.normalize_mcp_tools(), and passed to akasha.agents(tools=...). The supported transports are local stdio and remote Streamable HTTP. New integrations should use one Streamable HTTP /mcp endpoint; the older HTTP+SSE transport is deprecated.

The complete example is in examples/ex_mcp.py, with its server in examples/mcp_server.py. It uses tool_name_prefix=True when aggregating servers, preserves structured MCP results, and uses stream=False because MCP tools may be async-only.

For deterministic CI, use the local stdio fixture. Remote MCP tests must remain opt-in and should not require external credentials for the basic test suite.

Provider loading

Provider adapters are loaded when their provider is selected:

Model alias	Adapter
`openai:` / `azure:`	`langchain_openai`
`gemini:`	`langchain_google_genai`
`anthropic:`	`langchain_anthropic`
`ollama:`	`langchain_ollama`

Embedding adapters follow the same rule: the relevant embedding integration is loaded only when that embedding path is used. Common LangChain core modules are still shared by all providers.

Local development

Clone the repository and install it in editable mode:

git clone https://github.com/iii-org/akasha.git
cd akasha
uv venv --python 3.11

# Windows PowerShell
.venv\Scripts\Activate.ps1

uv pip install -e ".[light,dev]"

Run examples:

python examples/ex_ask.py
python examples/ex_rag.py
python examples/ex_agent.py

Testing

Unit tests do not require provider API keys:

python -m pytest tests -m unit

Focused agent and model tests:

python -m pytest \
  tests/provider/thinking/test_thinking_config.py \
  tests/agent/basic/test_core.py \
  tests/provider/factory/test_import_boundaries.py

Live provider tests are opt-in because they use API quota:

$env:RUN_LLM_TESTS = "1"
$env:ENV_FILE = "tests/.env"
python -m pytest tests/agent/stream/test_live_gemini.py -q

Live tests validate provider wiring, response types, tool calling, streaming events, and RAG flow. They do not evaluate the quality of model answers.

API overview

akasha.ask(...)           # document-aware QA and chat
akasha.agents(...)        # native tool-calling agent
akasha.RAG(...)           # document ingestion and retrieval
akasha.summary(...)       # map-reduce or refine summaries
akasha.MemoryManager(...) # persistent semantic memory

For detailed design decisions, upgrade notes, testing matrices, Skills, and runtime work, see dev_docs/.

License

Akasha is released under the MIT License.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

akasha_terminal-1.6.tar.gz (138.2 kB view details)

Uploaded Aug 13, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

akasha_terminal-1.6-py3-none-any.whl (175.4 kB view details)

Uploaded Aug 13, 2026 Python 3

File details

Details for the file akasha_terminal-1.6.tar.gz.

File metadata

Download URL: akasha_terminal-1.6.tar.gz
Upload date: Aug 13, 2026
Size: 138.2 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/7.0.0 CPython/3.11.15

File hashes

Hashes for akasha_terminal-1.6.tar.gz
Algorithm	Hash digest
SHA256	`1c3890f2be81543990e0673547914bb39941a435be2bff249a23fa9e4b6c0913`
MD5	`c9e4b78a9d70fa021b979f79d0b83d2c`
BLAKE2b-256	`f952fb6d995482fc5ecbb1fbf4aab20c6a2e598f48d9ff9604366265c2c33c14`

See more details on using hashes here.

File details

Details for the file akasha_terminal-1.6-py3-none-any.whl.

File metadata

Download URL: akasha_terminal-1.6-py3-none-any.whl
Upload date: Aug 13, 2026
Size: 175.4 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/7.0.0 CPython/3.11.15

File hashes

Hashes for akasha_terminal-1.6-py3-none-any.whl
Algorithm	Hash digest
SHA256	`e08e054809003e47f8754704fa8be1bcd3d8baae2e29489ea94a963212b909d3`
MD5	`22050f90671bb406bb5af8ea77b4447e`
BLAKE2b-256	`92a88a933dd85f6394267edec5cd1f7c551948bbcf04eacabefe6b74fc617b0b`