CBRkit

PyPI | Docs | Example

Customizable Case-Based Reasoning (CBR) toolkit for Python with a built-in API and CLI.

CBRkit Presentation
ICCBR 2024 Best Student Paper

---

CBRkit is a customizable and modular toolkit for Case-Based Reasoning (CBR) in Python. It provides a set of tools for loading cases and queries, defining similarity measures, and retrieving cases based on a query. The toolkit is designed to be flexible and extensible, allowing you to define custom similarity measures or use built-in ones. Retrieval pipelines are declared by composing these metrics, and the toolkit provides utility functions for applying them on a casebase. Additionally, it offers ready-to-use API and CLI interfaces for easy integration into your projects. The library is fully typed, enabling autocompletion and type checking in modern IDEs like VSCode and PyCharm.

To get started, we provide a demo project which contains a casebase and a predefined retriever. Further examples can be found in our tests and documentation. The following modules are part of CBRkit:

• cbrkit.loaders and cbrkit.dumpers: Functions for loading and exporting cases and queries.
• cbrkit.sim: Similarity generator functions for common data types like strings and numbers.
• cbrkit.retrieval: Functions for defining and applying retrieval pipelines.
• cbrkit.adapt: Adaptation generator functions for adapting cases based on a query.
• cbrkit.reuse: Functions for defining and applying reuse pipelines.
• cbrkit.typing: Generic type definitions for defining custom functions.
• cbrkit.synthesis: Functions for working on a casebase with LLMs to create new insights, e.g. in a RAG context.

Installation

The library is available on PyPI, so you can install it with pip:

pip install cbrkit

It comes with several optional dependencies for certain tasks like NLP which can be installed with:

pip install cbrkit[EXTRA_NAME,...]

where EXTRA_NAME is one of the following:

• all: All optional dependencies
• api: REST API Server
• cli: Command Line Interface (CLI)
• eval: Evaluation tools for common metrics like precision and recall
• graphs: Graph libraries like networkx and rustworkx
• llm: Large Language Models (LLM) APIs like Ollama and OpenAI
• nlp: Standalone NLP tools levenshtein, nltk, openai, and spacy
• timeseries: Time series similarity measures like dtw and smith_waterman
• transformers: Advanced NLP tools based on pytorch and transformers

Alternatively, you can also clone this git repository and install CBRKit and its dependencies via uv: uv sync --all-extras

Loading Cases

The first step is to load cases and queries. We provide predefined functions for the following formats:

• csv
• json
• toml
• xml
• yaml
• py (object inside of a python file).

Loading one of those formats can be done via the file function:

import cbrkit
casebase = cbrkit.loaders.file("path/to/cases.[json,toml,yaml,xml,csv]")

Additionally, CBRkit also integrates with polars and pandas for loading data frames. The following example shows how to load cases and queries from a CSV file using polars:

import polars as pl
import cbrkit

df = pl.read_csv("path/to/cases.csv")
casebase = cbrkit.loaders.polars(df)

Defining Queries

CBRkit expects the type of the queries to match the type of the cases. You may define a single query directly in Python as a dict:

query = {"name": "John", "age": 25}

If you have a collection of queries, you can load them using the same loader functions as for the cases.

 # for polars
queries = cbrkit.loaders.polars(pl.read_csv("path/to/queries.csv"))
# for any other supported file format
queries = cbrkit.loaders.file("path/to/queries.[json,toml,yaml,xml,csv]")

In case your query collection only contains a single entry, you can use the singleton function to extract it.

query = cbrkit.helpers.singleton(queries)

Similarity Measures and Aggregation

The next step is to define similarity measures for the cases and queries. It is possible to define custom measures, use built-in ones, or combine both.

Custom Similarity Measures

In CBRkit, a similarity measure is defined as a function that compares two arguments (a case and a query) and returns a similarity score: sim = f(x, y). It also supports pipeline-based similarity measures that are popular in NLP where a list of tuples is passed to the similarity measure: sims = f([(x1, y1), (x2, y2), ...]). This generic approach allows you to define custom similarity measures for your specific use case. For instance, the following function, which can be used to compare a string attribute of a case and a query, not only checks for strict equality, but also for partial matches (e.g., x = "blue" and y = "light blue"):

def color_similarity(x: str, y: str) -> float:
    if x == y:
        return 1.0
    elif x in y or y in x:
        return 0.5

    return 0.0

Please note: CBRkit inspects the signature of custom similarity functions to perform some checks. You need to make sure that the two parameters are named x and y, otherwise CBRkit will throw an error.

Built-in Similarity Measures

CBRkit also contains a selection of built-in similarity measures for the most common data types in the module cbrkit.sim. They are provided through generator functions that allow you to customize the behavior of the built-in measures. For example, an spacy-based embedding similarity measure can be obtained as follows:

semantic_similarity = cbrkit.sim.strings.spacy(model="en_core_web_lg")

Please note: Calling the function cbrkit.sim.strings.spacy returns a similarity function itself that has the same signature as the color_similarity function defined above.

An overview of all available similarity measures can be found in the module documentation.

Global Similarity and Aggregation

When dealing with cases that are not represented through elementary data types like strings, we need to aggregate individual measures to obtain a global similarity score. We provide a predefined aggregator that transforms a list of similarities into a single score. It can be used with custom and/or built-in measures.

similarities = [0.8, 0.6, 0.9]
aggregator = cbrkit.sim.aggregator(pooling="mean")
global_similarity = aggregator(similarities)

For the common use case of attribute-value based data, CBRkit provides a predefined global similarity measure that can be used as follows:

cbrkit.sim.attribute_value(
    attributes={
        "price": cbrkit.sim.numbers.linear(),
        "color": color_similarity # custom measure
        ...
    },
    aggregator=cbrkit.sim.aggregator(pooling="mean"),
)

The attribute_value function lets you define measures for each attribute of the cases/queries as well as the aggregation function. It also allows to use custom measures like the color_similarity function defined above.

Please note: The custom measure is not executed (i.e., there are no parenthesis at the end), but instead passed as a reference to the attribute_value function.

You may even nest similarity functions to create measures for object-oriented cases:

cbrkit.sim.attribute_value(
    attributes={
        "manufacturer": cbrkit.sim.attribute_value(
            attributes={
                "name": cbrkit.sim.strings.spacy(model="en_core_web_lg"),
                "country": cbrkit.sim.strings.levenshtein(),
            },
            aggregator=cbrkit.sim.aggregator(pooling="mean"),
        ),
        "color": color_similarity # custom measure
        ...
    },
    aggregator=cbrkit.sim.aggregator(pooling="mean"),
)

Retrieval

The final step is to retrieve cases based on the loaded queries. The cbrkit.retrieval module provides utility functions for this purpose. You first build a retrieval pipeline by specifying a global similarity function and optionally a limit for the number of retrieved cases.

retriever = cbrkit.retrieval.build(
    cbrkit.sim.attribute_value(...)
)

This retriever can then be applied on a casebase to retrieve cases for a given query.

result = cbrkit.retrieval.apply_query(casebase, query, retriever)

Our result has the following attributes:

• similarities: A dictionary containing the similarity scores for each case.
• ranking A list of case indices sorted by their similarity score.
• casebase The casebase containing only the retrieved cases (useful for downstream tasks).

An example using the provided cars-1k dataset can be found under examples/cars_retriever.py.

In some cases, it is useful to combine multiple retrieval pipelines, for example when applying the MAC/FAC pattern where a cheap pre-filter is applied to the whole casebase before a more expensive similarity measure is applied on the remaining cases. To use this pattern, first create the corresponding retrievers using the builder:

retriever1 = cbrkit.retrieval.dropout(..., min_similarity=0.5, limit=20)
retriever2 = cbrkit.retrieval.dropout(..., limit=10)

Then apply all of them sequentially by passing them as a list or tuple to the apply_query function:

result = cbrkit.retrieval.apply_query(casebase, query, (retriever1, retriever2))

The result has the following two attributes:

• final_step: Result of the last retriever in the list.
• steps: A list of results for each retriever in the list.

Both final_step and each entry in steps have the same attributes as discussed previously. The returned result also has these entries which are an alias for the corresponding entries in final_step (i.e., result.ranking == result.final_step.ranking).

Adaptation Functions

Case adaptation is a crucial step in the CBR cycle that allows us to modify retrieved cases to better suit the current query. CBRkit offers both built-in adaptation functions for common scenarios and the flexibility to define custom adaptation logic.

Please note: cbrkit.adapt contains the built-in adaption functions. To apply these (or custom adaption functions) to your actual casebase, please refer to Reuse.

Custom Adaptation Functions

In CBRkit, an adaptation function is defined as a function that takes two arguments (a case and a query) and returns an adapted case: adapted = f(case, query). For more complex scenarios, CBRkit also supports two additional types of adaptation functions:

• Map functions that operate on the entire casebase: adapted = f(casebase, query)
• Reduce functions that select and adapt a single case: key, adapted = f(casebase, query)

This generic approach allows you to define custom adaptation functions for your specific use case. For instance, the following function replaces a case value with the query value if they differ:

def replace_adapter(case: str, query: str) -> str:
    return query if case != query else case

Please note: CBRkit inspects the signature of custom adaptation functions to determine their type. Make sure that the parameters are named either case and query for pair functions, or casebase and query for map/reduce functions.

Built-in Adaptation Functions

CBRkit contains adaptation functions for common data types like strings and numbers in the module cbrkit.adapt. They are provided through generator functions that allow you to customize the behavior of the built-in functions. For example, a number aggregator can be obtained as follows:

# pooling must be a PoolingFunction or one of the provided PoolingNames
pooling = "mean"
number_adapter = cbrkit.adapt.numbers.aggregate(pooling)

Please note: Calling the function cbrkit.adapt.numbers.aggregate returns an adaptation function that takes a collection of values and returns an adapted value.

For the common use case of attribute-value based data, CBRkit provides a predefined global adapter that can be used as follows:

cbrkit.adapt.attribute_value(
    attributes={
        "price": cbrkit.adapt.numbers.aggregate(),
        "color": cbrkit.adapt.strings.regex("CASE_PATTERN", "QUERY_PATTERN", "REPLACEMENT"),
        ...
    }
)

The attribute_value function lets you define adaptation functions for each attribute of the cases. You may even nest adaptation functions to handle object-oriented cases.

An overview of all available adaptation functions can be found in the module documentation.

Reuse

The reuse phase applies adaptation functions to retrieved cases. The cbrkit.reuse module provides utility functions for this purpose. You first build a reuse pipeline by specifying a global adaptation function:

reuser = cbrkit.reuse.build(
    cbrkit.adapt.attribute_value(...),
)

This reuser can then be applied to the retrieval result to adapt cases based on a query:

result = cbrkit.reuse.apply_query(retrieval_result, query, reuser)

Our result has the following attributes:

• adaptations: A dictionary containing the adapted values for each case.
• ranking: A list of case indices matching the retrieval result.
• casebase: The casebase containing only the adapted cases.

Multiple reuse pipelines can be combined by passing them as a list or tuple:

reuser1 = cbrkit.reuse.build(...)
reuser2 = cbrkit.reuse.build(...)
result = cbrkit.reuse.apply_query(retrieval_result, query, (reuser1, reuser2))

The result structure follows the same pattern as the retrieval results with final_step and steps attributes.

Evaluation

CBRkit provides evaluation tools through the cbrkit.eval module for assessing the quality of retrieval results. The basic evaluation function cbrkit.eval.compute expects the following arguments:

• qrels: Ground truth relevance scores for query-case pairs. A higher value means a higher relevance.
• run: Retrieval similarity scores for query-case pairs.
• metrics: A list of metrics to compute.

You can evaluate retrieval results directly with the functions cbrkit.eval.retrieval and cbrkit.eval.retrieval_step.

Custom Metrics

Users can provide custom metric functions that implement the signature defined in the cbrkit.typing.EvalMetricFunc protocol:

def custom_metric(
    qrels: Mapping[str, Mapping[str, int]],
    run: Mapping[str, Mapping[str, float]],
    k: int,
    relevance_level: int,
) -> float:
    # Custom metric logic here
    return 0.0

You can then pass your custom metric function to the compute function:

results = cbrkit.eval.compute(
    qrels,
    run,
    metrics=["custom_metric@5"],
    metric_funcs={"custom_metric": custom_metric},
)

Built-in Metrics

The module also supports standard Information Retrieval metrics through ranx like precision, recall, and f1. A complete list is available in the ranx documentation. Additionally, CBRkit provides two custom metrics not available in ranx:

• correctness: Measures how well the ranking preserves the relevance ordering (-1 to 1).
• completeness: Measures what fraction of relevance pairs are preserved (0 to 1).

All of them can be computed at different cutoff points by appending @k, e.g., precision@5. We also offer a function to automatically generate a list of metrics for different cutoff points:

metrics = cbrkit.eval.metrics_at_k(["precision", "recall", "f1"], [1, 5, 10])

Synthesis

In the context of CBRkit, synthesis refers to creating new insights from the cases which were retrieved in a previous retrieval step, for example in a RAG context. CBRkit builds a synthesizer using the function cbrkit.synthesis.build with a provider and a prompt. A synthesizer maps a Result (obtained in the retrieval step) to an LLM output (can be a string or structurized). An example can be found in examples/cars_rag.py.

The following providers are currently supported if a valid API key is stored the respective environment variable:

• Anthropic (ANTHROPIC_API_KEY)
• Cohere (CO_API_KEY)
• Google (GOOGLE_API_KEY)
• Ollama
• OpenAI (OPENAI_API_KEY)

The respective provider class in cbrkit.synthesis.providers has to be initialized with the model name and a response type (either str or a Pydantic model for structured output). Further model options like temperature, seed, max_tokens, etc. can also be specified here.

A prompt produces an LLM input based on the specified instructions, an optional encoder (which maps a case or query to a string) and optional metadata. For a list of the currently included prompts, please refer to the module documentation

If the casebase is small enough, that it fits inside the LLM's context window, you can use CBRKit's synthesis as follows:

import cbrkit

casebase = cbrkit.loaders.LOADER(...)
retriever = cbrkit.retrieval.build(...)
retrieval = cbrkit.retrieval.apply_query(...)
provider = cbrkit.synthesis.providers.PROVIDER(...)
prompt = cbrkit.synthesis.prompts.default(instructions)
synthesizer = cbrkit.synthesis.build(provider, prompt)
response = cbrkit.synthesis.apply_result(retrieval, synthesizer).response

Working with large casebases

Because the built-in default and document_aware prompt functions include the entire casebase as context, the LLM input can be quite long when working with a large casebase. Because of this, in this case, we recommend transposing the cases (e.g. truncate every case to a fixed length) and/or apply chunking.

Transposing cases

CBRKit's transpose prompt allows to transpose cases and queries before they are passed to the main prompt function. This allows shortening entries like so:

from cbrkit.typing import JsonEntry
from cbrkit.dumpers import json_markdown

def encoder(value) -> dict:
    ...
baseprompt = cbrkit.synthesis.prompts.default(instructions, encoder=encoder)
# transform the entries, e.g. by shortening, leaving out irrelevant attributes, etc.
# In this case, the value of every field is trunctated to 100 characters
def shorten(entry: dict) -> JsonEntry:
    entry = {k: str(v)[:100] for k,v in entry.items()}
    return json_markdown(entry)

prompt = cbrkit.synthesis.prompts.transpose(baseprompt, shorten)
synthesizer = cbrkit.synthesis.build(provider, prompt)
...

Chunking

Instead of using cbrkit.synthesis.apply_result, CBRKit also provides the cbrkit.synthesis.chunks function to process the synthesis in batches. The partial results can then be aggregated using a pooling prompt.

import cbrkit

casebase = cbrkit.loaders.file(...)

queries = [...]
retriever = cbrkit.retrieval.dropout(...)
# run retrieval on each casebase chunk
retrievals = [cbrkit.retrieval.apply_query(casebase, query, retriever) for query in queries]

# batches are tuples of casebase, query, and retrieval similarities
batches = [(casebase, query, retrieval.similarities) for query, retrieval in zip(queries, retrievals)]

# Prompt which should be evaluated on each batch
prompt = cbrkit.synthesis.prompts.default(instructions="...")
provider = cbrkit.synthesis.providers.PROVIDER(...)
synthesizer = cbrkit.synthesis.build(provider, prompt)

# prompt to aggregate the partial results into a final result
pooling_prompt = cbrkit.synthesis.prompts.pooling(instructions="...")
pooling_func = cbrkit.synthesis.pooling(provider, pooling_prompt)
get_result = cbrkit.synthesis.chunks(synthesizer, pooling_func, size=10)
response = get_result(batches)

The complete version of this example can be found under examples/cars_rag_large.py.

Logging

CBRkit integrates with the logging module to provide a unified logging interface. By default, logging is not configured, you can activate by placing the following code in your project's __init__.py file:

import logging

logging.basicConfig(
    format="%(asctime)s %(levelname)s %(name)s: %(message)s",
    level=logging.INFO,
    datefmt="%Y-%m-%d %H:%M:%S",
)

logging.getLogger("cbrkit.sim.XXX").setLevel(logging.DEBUG) # to see debug messages from function XXX of the sim module
logging.getLogger("cbrkit.retrieval").setLevel(logging.DEBUG) # to see debug messages from the retrieval module

REST API and CLI

In order to use the built-in API and CLI, you need to define a retriever/reuser in a Python module using the function cbrkit.retrieval.build() and/or cbrkit.reuse.build(). For example, the file ./retriever_module.py could contain the following code:

import cbrkit

custom_retriever = cbrkit.retrieval.dropout(
    cbrkit.retrieval.build(
        cbrkit.sim.attribute_value(...),
    ),
    limit=10,
)

Our custom retriever can then be specified for the API/CLI using standard Python module syntax: retriever_module:custom_retriever.

CLI

When installing with the cli extra, CBRkit provides a command line interface:

cbrkit --help

Please visit the documentation for more information on how to use the CLI.

API

When installing with the api extra, CBRkit provides a REST API server:

cbrkit serve --help

After starting the server, you can access the API documentation at http://localhost:8080/docs.