Skip to content

function-synthesis.md

Latest commit

 

History

History
109 lines (81 loc) · 3.54 KB

function-synthesis.md

File metadata and controls

109 lines (81 loc) · 3.54 KB
Raw

Function Synthesis

Compile natural-language specs into callable neural artifacts (.chi bundles), run them locally, and expose them as agent tools.

Install

pip install 'chimera[function_synthesis]'

Pulls in llama-cpp-python (runtime) and huggingface_hub (base model cache).

60-second quickstart

from chimera.function_synthesis import FunctionSpec
from chimera.function_synthesis.compilers.mock import MockCompiler
from chimera.function_synthesis.registry import ProgramRegistry

spec = FunctionSpec(
    name="sentiment",
    description="Classify the text as 'positive' or 'negative'.",
)
bundle = MockCompiler().compile(spec)                     # no network, no training
slug = ProgramRegistry.default().install(spec=spec, bundle=bundle)
print(slug)   # -> "sentiment-ab12cd34"

CLI

# compile a spec file into a bundle + install it
echo '{"name":"sentiment","description":"classify pos/neg"}' > spec.json
chimera fs compile spec.json --compiler mock

# list installed programs
chimera fs list

# invoke one (requires a base GGUF + llama-cpp-python)
chimera fs run sentiment-ab12cd34 "I loved it!" \
    --base-model ~/.chimera/function_synthesis/models/.../model.gguf

# inspect / remove
chimera fs info sentiment-ab12cd34
chimera fs rm sentiment-ab12cd34

Using a real compile service

Point RemoteCompiler at any endpoint implementing the compile protocol:

from chimera.function_synthesis.compilers.remote import RemoteCompiler
from chimera.function_synthesis.strategies.synthesis import FunctionSynthesisStrategy

compiler = RemoteCompiler(endpoint="https://your-service/compile", api_key="...")
strategy = FunctionSynthesisStrategy(
    compiler=compiler,
    output_dir="./bundles",
)
result = strategy.run(spec)
print(result.bundle_path)

Agent tool

Any loaded CompiledFunction can be exposed to a Chimera agent:

from chimera.function_synthesis.backends.llama_cpp import LlamaCppBackend
from chimera.function_synthesis.runtime import CompiledFunction
from chimera.tools.compiled_function_tool import CompiledFunctionTool

backend = LlamaCppBackend(base_model_path="...")
fn = CompiledFunction.from_path("sentiment-ab12cd34.chi", backend=backend)
tool = CompiledFunctionTool(fn)

# add `tool` to your agent's tool list --- agents can now call the compiled
# function just like any other tool.

Cache layout

$CHIMERA_FS_HOME/  (default: ~/.chimera/function_synthesis/)
  models/       # base GGUF files, downloaded on first use
  bundles/      # installed .chi bundles (one per slug)
  index.json    # slug -> bundle path + metadata
  prefix/       # cold-start state cache (see PrefixCache)

Set CHIMERA_FS_HOME to override. Set CHIMERA_FS_OFFLINE=1 to refuse any network call --- cache misses raise OfflineError instead of downloading.

.chi bundle format

A .chi file is a ZIP archive containing:

Member Contents
manifest.json Schema version, name, base model ID, adapter format
adapter.gguf LoRA adapter (Q4_0 by default)
prompts.json {"system": str, "user_template": str, "stop": [str]}
spec.json Serialized FunctionSpec
metadata.json Compiler backend info, base model hash, free-form fields

See chimera/function_synthesis/bundle.py for the loader.