Langda

Language-Driven Agent for Probabilistic Logic Programming

Automatically generate ProbLog code from natural language using LLM agents.

Installation

# From GitHub
pip install git+https://github.com/Symbolic-Intelligence-Org/langda-project.git

# Or clone and install locally
git clone https://github.com/Symbolic-Intelligence-Org/langda-project.git
cd langda-project
pip install -e .

To enable LangChain logging integration, you need to install langchain-logger manually:

pip install --no-deps langchain-logger==0.1.0

!We use --no-deps here to avoid pulling in incompatible dependencies.

To enable retrieve function, you could also install faiss cpu or faiss gpu, this is optional.

Update

pip install --upgrade git+https://github.com/Symbolic-Intelligence-Org/langda-project.git

If you feel you are unable to update, use:

pip install --upgrade --force-reinstall --no-cache-dir git+https://github.com/Symbolic-Intelligence-Org/langda-project.git

Quick Start

from langda import langda_solve

rules = """
langda(LLM:"Define factorial predicate").
query(factorial(5, X)).
"""

result = langda_solve(
    agent_type="double_dc",
    rule_string=rules,
    model_name="deepseek-chat"
)

print(result)

langda_solve — Unified Entry for LangDa Execution

This is the central API for executing LangDa workflows. It dynamically selects the appropriate agent architecture and runs the full generation–evaluation–refinement process.

Function Signature

langda_solve(
    rule_string: str,
    **overrides: Unpack[SolveOverrides]
) -> str

Parameters

rule_string (str, required)
: ProbLog or hybrid LangDa rules to process. Must be provided.

agent_type (Literal["single_simple","double_simple","single_dc","double_dc"], default="single_dc")
: Select the agent architecture:
– single_*: generate-only
– double_*: generate–evaluate–refine
– _simple: simple agent
– _dc: double-chain (recommended)

model_name (str, default="deepseek-chat")
: The model name used by your API key.

prefix (str, default="")
: Optional prefix to differentiate output files or database entries.

save_dir (str | Path, default=current directory)
: Folder for outputs and cached results.

load (bool, default=False)
: If True, directly load from database, skipping generation when available.

langda_ext (dict, default={})
: Dynamic content mapping for placeholders.
Example: langda(LLM:"/* City */ weather") → {"City": "Berlin"}

query_ext (str, default="")
: For DeepProbLog tasks, add extra facts or queries if needed.

log_path (str, default="langda.log")
: Log file name; combined with prefix if set.

config (dict, optional)
: Optional session configuration.

api_key (str, optional)
: Optional override for model API key.

Default Configuration Example

config = {
    "configurable": {
        "thread_id": str(uuid4()),
        "checkpoint_ns": "langda",
        "checkpoint_id": None,
    },
    "metadata": {}
}

Agent Map

Return

str — The final executable code or result from the LangDa workflow.

Example

rules = """
% Simple example
langda(LLM:"Describe today's weather in Paris", LOT:"search").
weather(paris, sunny, 25).
"""

result = langda_solve(
    rule_string=rules,
    agent_type="double_dc",
    model_name="deepseek-chat",
    prefix="weather_demo",
    save_dir="./outputs",
    log_path="demo.log"
)
print(result)

Notes

• langda_solve automatically sets up logging and prints start/finish markers for each run.
• The double-chain agent (double_dc) is the most capable and recommended mode.

Configuration

Create .env file:

# DeepSeek (recommended)
GNRT_DEEPSEEK_PROVIDER=deepseek
GNRT_DEEPSEEK_MODEL=deepseek-chat
GNRT_DEEPSEEK_API_KEY=your-api-key
GNRT_DEEPSEEK_API_TYP=Bearer
GNRT_DEEPSEEK_API_VER=2025-03-15

# Optional: for web search function
TAVILY_API_KEY=your-tavily-api-key

For OpenAI or Groq, replace DEEPSEEK with OPENAI or GROQ.

Agent Types

• single_simple - Basic generation
• double_simple - Generation with evaluation
• single_dc - Dual-phase generation ⭐ (recommended)
• double_dc - Dual-phase with evaluation

Langda Syntax

LangDa introduces a unified predicate langda/3 as the central interface between natural language and probabilistic logic programming.
It allows users to describe rules, facts, or reasoning steps directly in English while maintaining full ProbLog compatibility.

Each langda predicate can include up to three parameters:

langda(LLM:"<instruction>", LOT:"<tool>", FUP:"<policy>").

• LLM – (required) natural language instruction or description.
• LOT – (optional) external tool specification ("search" for web, "retrieve" for local DB`).
• FUP – (optional) forced update flag ("true" regenerates every run, "false" only when changed`).

Example:

langda(LLM:"What was the weather yesterday in Darmstadt?", LOT:"search", FUP:"true").

LangDa automatically interprets linguistic uncertainty (e.g., often, rarely) into probabilistic annotations
and integrates surrounding facts to infer variable bindings, generating executable ProbLog clauses.

Examples

Dynamic Content

rules = """
langda(LLM:"Define rules for /* City */").
"""

result = langda_solve(
    rule_string=rules,
    agent_type="double_dc",
    model_name="deepseek-chat",
    langda_ext={"City": "Tokyo"}
)
print(result)

EXT Usage (langda_ext / query_ext)

from langda import langda_solve

# 1) Use langda_ext to inject dynamic placeholders into LLM prompts
rules_dynamic = r"""
% The placeholders /* City */ and /* Task */ will be replaced by langda_ext
langda(LLM:"Create /* Task */ rules for /* City */, include base facts and a query example.").
"""

result_dynamic = langda_solve(
    rule_string=rules_dynamic,
    agent_type="double_dc",
    model_name="deepseek-chat",
    prefix="dynamic_ext_demo",
    save_dir="./outputs",
    langda_ext={
        "City": "Berlin",
        "Task": "weather"
    },
    log_path="dynamic_ext.log"
)
print(result_dynamic)

# 2) For DeepProbLog, use query_ext to append extra facts/queries at the end
rules_dpl = r"""
% Generate a probabilistic model and leave space for external queries
langda(LLM:"Define a simple coin model with probabilities and an observation.").
"""

extra_queries = r"""
% --- query_ext appended content ---
evidence(coin, heads).
query(coin).
"""

result_dpl = langda_solve(
    rule_string=rules_dpl,
    agent_type="double_dc",
    model_name="deepseek-chat",
    prefix="deepproblog_ext_demo",
    save_dir="./outputs",
    query_ext=extra_queries,
    log_path="deepproblog_ext.log"
)
print(result_dpl)

Knowledge Base (Optional)

For retriever tool, create langda/utils/problog_docs.json:

[
  {
    "id": "example_1",
    "title": "Title",
    "content": "Content here...",
    "tags": ["tag1"],
    "keywords": ["keyword1", "keyword2"]
  }
]