linkml-store 0.2.0__tar.gz → 0.2.1__tar.gz

{linkml_store-0.2.0 → linkml_store-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: linkml-store
-Version: 0.2.0
+Version: 0.2.1
 Summary: linkml-store
 License: MIT
 Author: Author 1
@@ -18,6 +18,7 @@ Provides-Extra: chromadb
 Provides-Extra: fastapi
 Provides-Extra: frictionless
 Provides-Extra: h5py
+Provides-Extra: ibis
 Provides-Extra: llm
 Provides-Extra: map
 Provides-Extra: mongodb
@@ -34,7 +35,9 @@ Requires-Dist: duckdb (>=0.10.1)
 Requires-Dist: duckdb-engine (>=0.11.2)
 Requires-Dist: fastapi ; extra == "fastapi"
 Requires-Dist: frictionless ; extra == "frictionless"
+Requires-Dist: gcsfs ; extra == "ibis"
 Requires-Dist: h5py ; extra == "h5py"
+Requires-Dist: ibis-framework[duckdb,examples] (>=9.3.0) ; extra == "ibis"
 Requires-Dist: jinja2 (>=3.1.4,<4.0.0)
 Requires-Dist: jsonlines (>=4.0.0,<5.0.0)
 Requires-Dist: linkml (>=1.8.0) ; extra == "validation"
@@ -43,6 +46,7 @@ Requires-Dist: linkml_map ; extra == "map"
 Requires-Dist: linkml_renderer ; extra == "renderer"
 Requires-Dist: llm ; extra == "llm"
 Requires-Dist: matplotlib ; extra == "analytics"
+Requires-Dist: multipledispatch ; extra == "ibis"
 Requires-Dist: neo4j ; extra == "neo4j"
 Requires-Dist: networkx ; extra == "neo4j"
 Requires-Dist: pandas (>=2.2.1) ; extra == "analytics"
@@ -52,6 +56,7 @@ Requires-Dist: pyarrow ; extra == "pyarrow"
 Requires-Dist: pydantic (>=2.0.0,<3.0.0)
 Requires-Dist: pymongo ; extra == "mongodb"
 Requires-Dist: pystow (>=0.5.4,<0.6.0)
+Requires-Dist: ruff (>=0.6.2) ; extra == "tests"
 Requires-Dist: scikit-learn ; extra == "scipy"
 Requires-Dist: scipy ; extra == "scipy"
 Requires-Dist: seaborn ; extra == "analytics"

{linkml_store-0.2.0 → linkml_store-0.2.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "linkml-store"
-version = "0.2.0"
+version = "0.2.1"
 description = "linkml-store"
 authors = ["Author 1 <author@org.org>"]
 license = "MIT"
@@ -20,6 +20,7 @@ seaborn = { version = "*", optional = true }
 plotly = { version = "*", optional = true }
 pystow = "^0.5.4"
 black = { version=">=24.0.0", optional = true }
+ruff = { version=">=0.6.2", optional = true }
 llm = { version="*", optional = true }
 tiktoken = { version="*", optional = true }
 pymongo = { version="*", optional = true }
@@ -35,6 +36,9 @@ linkml = { version=">=1.8.0", optional = true }
 linkml_map = { version="*", optional = true }
 linkml_renderer = { version="*", optional = true }
 frictionless = { version="*", optional = true }
+ibis-framework = { version=">=9.3.0", extras = ["duckdb", "examples"], optional = true }
+gcsfs = { version="*", optional = true }
+multipledispatch = { version="*" }
 pandas = ">=2.2.1"
 jinja2 = "^3.1.4"
 jsonlines = "^4.0.0"
@@ -69,7 +73,7 @@ numpy = [
 [tool.poetry.extras]
 analytics = ["pandas", "matplotlib", "seaborn", "plotly"]
 app = ["streamlit"]
-tests = ["black"]
+tests = ["black", "ruff"]
 llm = ["llm", "tiktoken"]
 mongodb = ["pymongo"]
 neo4j = ["neo4j", "py2neo", "networkx"]
@@ -82,6 +86,7 @@ renderer = ["linkml_renderer"]
 fastapi = ["fastapi", "uvicorn"]
 frictionless = ["frictionless"]
 scipy = ["scipy", "scikit-learn"]
+ibis = ["ibis-framework", "multipledispatch", "gcsfs"]
 
 [tool.poetry.scripts]
 linkml-store = "linkml_store.cli:cli"
@@ -119,27 +124,13 @@ extend-exclude = [
 ]
 force-exclude = true
 line-length = 120
-extend-ignore = ["E203"]
-select = [
+lint.extend-ignore = ["E203"]
+lint.select = [
   "E",  # pycodestyle errors
   "F",  # Pyflakes
   "I",  # isort
 ]
-# Assume Python 3.8
-target-version = "py38"
 
-[tool.ruff.per-file-ignores]
-# These templates can have long lines
-"linkml/generators/sqlalchemy/sqlalchemy_declarative_template.py" = ["E501"]
-"linkml/generators/sqlalchemy/sqlalchemy_imperative_template.py" = ["E501"]
-
-# Notebooks can have unsorted imports
-"tests/test_notebooks/input/*" = ["E402"]
-
-
-[tool.ruff.mccabe]
-# Unlike Flake8, default to a complexity level of 10.
-max-complexity = 10
 
 
 [tool.codespell]

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/api/collection.py

@@ -226,6 +226,18 @@ class Collection(Generic[DatabaseType]):
         self._initialized = True
         patches = [{"op": "add", "path": "/0", "value": obj} for obj in objs]
         self._broadcast(patches, **kwargs)
+        self._post_modification_hook(**kwargs)
+
+    def _post_delete_hook(self, **kwargs):
+        self._post_modification_hook(**kwargs)
+
+    def _post_modification_hook(self, **kwargs):
+        for indexer in self.indexers.values():
+            ix_collection_name = self.get_index_collection_name(indexer)
+            ix_collection = self.parent.get_collection(ix_collection_name)
+            # Currently updating the source triggers complete reindexing
+            # TODO: make this more efficient by only deleting modified
+            ix_collection.delete_where({})
 
     def delete(self, objs: Union[OBJECT, List[OBJECT]], **kwargs) -> Optional[int]:
         """
@@ -476,7 +488,7 @@ class Collection(Generic[DatabaseType]):
         Now let's index, using the simple trigram-based index
 
         >>> index = get_indexer("simple")
-        >>> collection.attach_indexer(index)
+        >>> _ = collection.attach_indexer(index)
 
         Now let's find all objects:
 
@@ -514,7 +526,10 @@ class Collection(Generic[DatabaseType]):
         if ix_coll.size() == 0:
             logger.info(f"Index {index_name} is empty; indexing all objects")
             all_objs = self.find(limit=-1).rows
-            self.index_objects(all_objs, index_name, replace=True, **kwargs)
+            if all_objs:
+                # print(f"Index {index_name} is empty; indexing all objects {len(all_objs)}")
+                self.index_objects(all_objs, index_name, replace=True, **kwargs)
+                assert ix_coll.size() > 0
         qr = ix_coll.find(where=where, limit=-1, **kwargs)
         index_col = ix.index_field
         # TODO: optimize this for large indexes
@@ -648,7 +663,31 @@ class Collection(Generic[DatabaseType]):
         """
         return self.find({}, limit=1).num_rows
 
-    def attach_indexer(self, index: Union[Indexer, str], name: Optional[str] = None, auto_index=True, **kwargs):
+    def rows_iter(self) -> Iterable[OBJECT]:
+        """
+        Return an iterator over the objects in the collection.
+
+        :return:
+        """
+        yield from self.find({}, limit=-1).rows
+
+    def rows(self) -> List[OBJECT]:
+        """
+        Return a list of objects in the collection.
+
+        :return:
+        """
+        return list(self.rows_iter())
+
+    def ranked_rows(self) -> List[Tuple[float, OBJECT]]:
+        """
+        Return a list of objects in the collection, with scores.
+        """
+        return [(n, obj) for n, obj in enumerate(self.rows_iter())]
+
+    def attach_indexer(
+        self, index: Union[Indexer, str], name: Optional[str] = None, auto_index=True, **kwargs
+    ) -> Indexer:
         """
         Attach an index to the collection.
 
@@ -669,8 +708,8 @@ class Collection(Generic[DatabaseType]):
         >>> full_index.name = "full"
         >>> name_index = get_indexer("simple", text_template="{name}")
         >>> name_index.name = "name"
-        >>> collection.attach_indexer(full_index)
-        >>> collection.attach_indexer(name_index)
+        >>> _ = collection.attach_indexer(full_index)
+        >>> _ = collection.attach_indexer(name_index)
 
         Now let's find objects using the full index, using the string "France".
         We expect the country France to be the top hit, but the score will
@@ -713,6 +752,10 @@ class Collection(Generic[DatabaseType]):
             all_objs = self.find(limit=-1).rows
             logger.info(f"Auto-indexing {len(all_objs)} objects")
             self.index_objects(all_objs, index_name, replace=True, **kwargs)
+        return index
+
+    def get_index_collection_name(self, indexer: Indexer) -> str:
+        return self._index_collection_name(indexer.name)
 
     def _index_collection_name(self, index_name: str) -> str:
         """

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/api/database.py

@@ -268,7 +268,7 @@ class Database(ABC, Generic[CollectionType]):
         metadata: Optional[CollectionConfig] = None,
         recreate_if_exists=False,
         **kwargs,
-    ) -> CollectionType:
+    ) -> Collection:
         """
         Create a new collection in the current database.
 
@@ -760,6 +760,12 @@ class Database(ABC, Generic[CollectionType]):
         """
         Export a database to a file or location.
 
+        >>> from linkml_store.api.client import Client
+        >>> client = Client()
+        >>> db = client.attach_database("duckdb", alias="test")
+        >>> db.import_database("tests/input/iris.csv", Format.CSV, collection_name="iris")
+        >>> db.export_database("/tmp/iris.yaml", Format.YAML)
+
         :param location: location of the file
         :param target_format: target format
         :param kwargs: additional arguments

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/api/queries.py

@@ -40,7 +40,9 @@ class FacetCountResult(BaseModel):
 
 class QueryResult(BaseModel):
     """
-    A query result
+    A query result.
+
+    TODO: make this a subclass of Collection
     """
 
     query: Optional[Query] = None

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/api/stores/duckdb/duckdb_collection.py

@@ -50,8 +50,9 @@ class DuckDBCollection(Collection):
         if not isinstance(objs, list):
             objs = [objs]
         cd = self.class_definition()
-        if not cd:
+        if not cd or not cd.attributes:
             cd = self.induce_class_definition_from_objects(objs)
+        assert cd.attributes
         table = self._sqla_table(cd)
         engine = self.parent.engine
         with engine.connect() as conn:
@@ -61,7 +62,8 @@ class DuckDBCollection(Collection):
                 stmt = stmt.compile(engine)
                 conn.execute(stmt)
                 conn.commit()
-        return
+        self._post_delete_hook()
+        return None
 
     def delete_where(self, where: Optional[Dict[str, Any]] = None, missing_ok=True, **kwargs) -> Optional[int]:
         logger.info(f"Deleting from {self.target_class_name} where: {where}")
@@ -87,6 +89,7 @@ class DuckDBCollection(Collection):
             if deleted_rows_count == 0 and not missing_ok:
                 raise ValueError(f"No rows found for {where}")
             conn.commit()
+            self._post_delete_hook()
             return deleted_rows_count if deleted_rows_count > -1 else None
 
     def query_facets(

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/cli.py

@@ -76,6 +76,8 @@ class ContextSettings(BaseModel):
         if name is None:
             # if len(self.database.list_collections()) > 1:
             #    raise ValueError("Collection must be specified if there are multiple collections.")
+            if not self.database:
+                return None
             if not self.database.list_collections():
                 return None
             name = list(self.database.list_collections())[0]
@@ -218,7 +220,10 @@ def insert(ctx, files, object, format):
 @click.option("--object", "-i", multiple=True, help="Input object as YAML")
 @click.pass_context
 def store(ctx, files, object, format):
-    """Store objects from files (JSON, YAML, TSV) into the specified collection."""
+    """Store objects from files (JSON, YAML, TSV) into the database.
+
+    Note: this is similar to insert, but a collection does not need to be specified
+    """
     settings = ctx.obj["settings"]
     db = settings.database
     if not files and not object:
@@ -499,6 +504,7 @@ def describe(ctx, where, output_type, output, limit):
     "--predictor-type", "-t", default="sklearn", show_default=True, type=click.STRING, help="Type of predictor"
 )
 @click.option("--evaluation-count", "-n", type=click.INT, help="Number of examples to evaluate over")
+@click.option("--evaluation-match-function", help="Name of function to use for matching objects in eval")
 @click.option("--query", "-q", type=click.STRING, help="query term")
 @click.pass_context
 def infer(
@@ -506,6 +512,7 @@ def infer(
     inference_config_file,
     query,
     evaluation_count,
+    evaluation_match_function,
     training_test_data_split,
     predictor_type,
     target_attribute,
@@ -549,7 +556,10 @@ def infer(
     else:
         query_obj = None
     collection = ctx.obj["settings"].collection
-    atts = collection.class_definition().attributes.keys()
+    if collection:
+        atts = collection.class_definition().attributes.keys()
+    else:
+        atts = []
     if feature_attributes:
         features = feature_attributes.split(",")
         features = [f.strip() for f in features]
@@ -575,7 +585,8 @@ def infer(
         if training_test_data_split:
             config.train_test_split = training_test_data_split
         predictor = get_inference_engine(predictor_type, config=config)
-        predictor.load_and_split_data(collection)
+        if collection:
+            predictor.load_and_split_data(collection)
         predictor.initialize_model()
     if export_model:
         logger.info(f"Exporting model to {export_model} in {model_format}")
@@ -584,8 +595,14 @@ def infer(
         if not export_model and not evaluation_count:
             raise ValueError("Query or evaluate must be specified if not exporting model")
     if evaluation_count:
+        if evaluation_match_function == "score_text_overlap":
+            match_function_fn = score_text_overlap
+        elif evaluation_match_function is not None:
+            raise ValueError(f"Unknown match function: {evaluation_match_function}")
+        else:
+            match_function_fn = None
         outcome = evaluate_predictor(
-            predictor, target_attributes, evaluation_count=evaluation_count, match_function=score_text_overlap
+            predictor, target_attributes, evaluation_count=evaluation_count, match_function=match_function_fn
         )
         print(f"Outcome: {outcome} // accuracy: {outcome.accuracy}")
     if query_obj:

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/index/implementations/llm_indexer.py

@@ -1,11 +1,13 @@
 import logging
 from pathlib import Path
-from typing import TYPE_CHECKING, List
+from typing import TYPE_CHECKING, List, Optional
 
 import numpy as np
+from tiktoken import encoding_for_model
 
 from linkml_store.api.config import CollectionConfig
 from linkml_store.index.indexer import INDEX_ITEM, Indexer
+from linkml_store.utils.llm_utils import get_token_limit, render_formatted_text
 
 if TYPE_CHECKING:
     import llm
@@ -29,6 +31,7 @@ class LLMIndexer(Indexer):
     cached_embeddings_database: str = None
     cached_embeddings_collection: str = None
     cache_queries: bool = False
+    truncation_method: Optional[str] = None
 
     @property
     def embedding_model(self):
@@ -62,6 +65,21 @@ class LLMIndexer(Indexer):
         """
         logging.info(f"Converting {len(texts)} texts to vectors")
         model = self.embedding_model
+        token_limit = get_token_limit(model.model_id)
+        encoding = encoding_for_model("gpt-4o")
+
+        def truncate_text(text: str) -> str:
+            # split into tokens every 1000 chars:
+            parts = [text[i : i + 1000] for i in range(0, len(text), 1000)]
+            return render_formatted_text(
+                lambda x: "".join(x),
+                parts,
+                encoding,
+                token_limit,
+            )
+
+        texts = [truncate_text(text) for text in texts]
+
         if self.cached_embeddings_database and (cache is None or cache or self.cache_queries):
             model_id = model.model_id
             if not model_id:
@@ -88,7 +106,7 @@ class LLMIndexer(Indexer):
                 embeddings_collection = embeddings_db.create_collection(coll_name, metadata=config)
             else:
                 embeddings_collection = embeddings_db.create_collection(coll_name, metadata=config)
-            texts = list(texts)
+
             embeddings = list([None] * len(texts))
             uncached_texts = []
             n = 0

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/index/indexer.py

@@ -36,6 +36,54 @@ def cosine_similarity(vector1, vector2) -> float:
 class Indexer(BaseModel):
     """
     An indexer operates on a collection in order to search for objects.
+
+    You should use a subcllass of this; this can be looked up dynqamically:
+
+    >>> from linkml_store.index import get_indexer
+    >>> indexer = get_indexer("simple")
+
+    You can customize how objects are indexed by passing in a text template.
+    For example, if your collection has objects with "name" and "profession" attributes,
+    you can index them as "{name} {profession}".
+
+    >>> indexer = get_indexer("simple", text_template="{name} :: {profession}")
+
+    By default, python fstrings are assumed.
+
+    We can test this works using the :ref:`object_to_text` method (normally
+    you would never need to call this directly, but it's useful for testing):
+
+    >>> obj = {"name": "John", "profession": "doctor"}
+    >>> indexer.object_to_text(obj)
+    'John :: doctor'
+
+    You can also use Jinja2 templates; this gives more flexibility and logic,
+    e.g. conditional formatting:
+
+    >>> tmpl = "{{name}}{% if profession %} :: {{profession}}{% endif %}"
+    >>> indexer = get_indexer("simple", text_template=tmpl, text_template_syntax=TemplateSyntaxEnum.jinja2)
+    >>> indexer.object_to_text(obj)
+    'John :: doctor'
+    >>> indexer.object_to_text({"name": "John"})
+    'John'
+
+    You can also specify which attributes to index:
+
+    >>> indexer = get_indexer("simple", index_attributes=["name"])
+    >>> indexer.object_to_text(obj)
+    'John'
+
+    The purpose of an indexer is to translate a collection of objects into a collection of objects
+    such as vectors for purposes such as search. Unless you are implementing your own indexer, you
+    generally don't need to use the methods that return vectors, but we can examine their behavior
+    to get a sense of how they work.
+
+    >>> vectors = indexer.objects_to_vectors([{"name": "Aardvark"}, {"name": "Aardwolf"}, {"name": "Zesty"}])
+    >>> assert cosine_similarity(vectors[0], vectors[1]) > cosine_similarity(vectors[0], vectors[2])
+
+    Note you should consult the documentation for the specific indexer you are using for more details on
+    how text is converted to vectors.
+
     """
 
     name: Optional[str] = None
@@ -122,7 +170,9 @@ class Indexer(BaseModel):
         self, query: str, vectors: List[Tuple[str, INDEX_ITEM]], limit: Optional[int] = None
     ) -> List[Tuple[float, Any]]:
         """
-        Search the index for a query string
+        Use the indexer to search against a database of vectors.
+
+        Note: this is a low-level method, typically you would use the :ref:`search` method on a :ref:`Collection`.
 
         :param query: The query string to search for
         :param vectors: A list of indexed items, where each item is a tuple of (id, vector)

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/inference/evaluation.py

@@ -20,6 +20,8 @@ def score_match(target: Optional[Any], candidate: Optional[Any], match_function:
     1.0
     >>> score_match("a", "b")
     0.0
+    >>> score_match("abcd", "abcde")
+    0.0
     >>> score_match("a", None)
     0.0
     >>> score_match(None, "a")
@@ -52,7 +54,7 @@ def score_match(target: Optional[Any], candidate: Optional[Any], match_function:
 
     :param target:
     :param candidate:
-    :param match_function:
+    :param match_function: defaults to struct
     :return:
     """
     if target == candidate:
@@ -99,7 +101,8 @@ def evaluate_predictor(
     :param predictor:
     :param target_attributes:
     :param feature_attributes:
-    :param evaluation_count:
+    :param evaluation_count: max iterations
+    :param match_function: function to use for matching
     :return:
     """
     n = 0
@@ -113,8 +116,8 @@ def evaluate_predictor(
         else:
             test_obj = row
         result = predictor.derive(test_obj)
-        logger.info(f"Predicted: {result.predicted_object} Expected: {expected_obj}")
         tp += score_match(result.predicted_object, expected_obj, match_function)
+        logger.info(f"TP={tp} MF={match_function} Predicted: {result.predicted_object} Expected: {expected_obj}")
         n += 1
         if evaluation_count is not None and n >= evaluation_count:
             break
@@ -125,6 +128,9 @@ def score_text_overlap(str1: Any, str2: Any) -> float:
     """
     Compute the overlap score between two strings.
 
+    >>> score_text_overlap("abc", "bcde")
+    0.5
+
     :param str1:
     :param str2:
     :return:

linkml_store-0.2.1/src/linkml_store/inference/implementations/rag_inference_engine.py

@@ -0,0 +1,232 @@
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import ClassVar, List, Optional, TextIO, Union
+
+import yaml
+from llm import get_key
+from pydantic import BaseModel
+
+from linkml_store.api.collection import OBJECT, Collection
+from linkml_store.inference.inference_config import Inference, InferenceConfig, LLMConfig
+from linkml_store.inference.inference_engine import InferenceEngine, ModelSerialization
+from linkml_store.utils.object_utils import select_nested
+
+logger = logging.getLogger(__name__)
+
+SYSTEM_PROMPT = """
+You are a {llm_config.role}, your task is to inference the YAML
+object output given the YAML object input. I will provide you
+with a collection of examples that will provide guidance both
+on the desired structure of the response, as well as the kind
+of content.
+
+You should return ONLY valid YAML in your response.
+"""
+
+
+class TrainedModel(BaseModel, extra="forbid"):
+    rag_collection_rows: List[OBJECT]
+    index_rows: List[OBJECT]
+    config: Optional[InferenceConfig] = None
+
+
+@dataclass
+class RAGInferenceEngine(InferenceEngine):
+    """
+    AI Retrieval Augmented Generation (RAG) based predictor.
+
+
+    >>> from linkml_store.api.client import Client
+    >>> from linkml_store.utils.format_utils import Format
+    >>> from linkml_store.inference.inference_config import LLMConfig
+    >>> client = Client()
+    >>> db = client.attach_database("duckdb", alias="test")
+    >>> db.import_database("tests/input/countries/countries.jsonl", Format.JSONL, collection_name="countries")
+    >>> db.list_collection_names()
+    ['countries']
+    >>> collection = db.get_collection("countries")
+    >>> features = ["name"]
+    >>> targets = ["code", "capital", "continent", "languages"]
+    >>> llm_config = LLMConfig(model_name="gpt-4o-mini",)
+    >>> config = InferenceConfig(target_attributes=targets, feature_attributes=features, llm_config=llm_config)
+    >>> ie = RAGInferenceEngine(config=config)
+    >>> ie.load_and_split_data(collection)
+    >>> ie.initialize_model()
+    >>> prediction = ie.derive({"name": "Uruguay"})
+    >>> prediction.predicted_object
+    {'capital': 'Montevideo', 'code': 'UY', 'continent': 'South America', 'languages': ['Spanish']}
+
+    The "model" can be saved for later use:
+
+    >>> ie.export_model("tests/output/countries.rag_model.json")
+
+    Note in this case the model is not the underlying LLM, but the "RAG Model" which is the vectorized
+    representation of training set objects.
+
+    """
+
+    _model: "llm.Model" = None  # noqa: F821
+
+    rag_collection: Collection = None
+
+    PERSIST_COLS: ClassVar[List[str]] = [
+        "config",
+    ]
+
+    def __post_init__(self):
+        if not self.config:
+            self.config = InferenceConfig()
+        if not self.config.llm_config:
+            self.config.llm_config = LLMConfig()
+
+    @property
+    def model(self) -> "llm.Model":  # noqa: F821
+        import llm
+
+        if self._model is None:
+            self._model = llm.get_model(self.config.llm_config.model_name)
+            if self._model.needs_key:
+                key = get_key(None, key_alias=self._model.needs_key)
+                self._model.key = key
+
+        return self._model
+
+    def initialize_model(self, **kwargs):
+        logger.info(f"Initializing model {self.model}")
+        if self.training_data:
+            rag_collection = self.training_data.collection
+            rag_collection.attach_indexer("llm", auto_index=False)
+            self.rag_collection = rag_collection
+
+    def object_to_text(self, object: OBJECT) -> str:
+        return yaml.dump(object)
+
+    def derive(self, object: OBJECT) -> Optional[Inference]:
+        import llm
+        from tiktoken import encoding_for_model
+
+        from linkml_store.utils.llm_utils import get_token_limit, render_formatted_text
+
+        model: llm.Model = self.model
+        model_name = self.config.llm_config.model_name
+        feature_attributes = self.config.feature_attributes
+        target_attributes = self.config.target_attributes
+        num_examples = self.config.llm_config.number_of_few_shot_examples or 5
+        query_text = self.object_to_text(object)
+        if not self.rag_collection:
+            # TODO: zero-shot mode
+            examples = []
+        else:
+            if not self.rag_collection.indexers:
+                raise ValueError("RAG collection must have an indexer attached")
+            rs = self.rag_collection.search(query_text, limit=num_examples, index_name="llm")
+            examples = rs.rows
+            if not examples:
+                raise ValueError(f"No examples found for {query_text}; size = {self.rag_collection.size()}")
+        prompt_clauses = []
+        query_obj = select_nested(object, feature_attributes)
+        query_text = self.object_to_text(query_obj)
+        for example in examples:
+            input_obj = select_nested(example, feature_attributes)
+            input_obj_text = self.object_to_text(input_obj)
+            if input_obj_text == query_text:
+                raise ValueError(
+                    f"Query object {query_text} is the same as example object {input_obj_text}\n"
+                    "This indicates possible test data leakage\n."
+                    "TODO: allow an option that allows user to treat this as a basic lookup\n"
+                )
+            output_obj = select_nested(example, target_attributes)
+            prompt_clause = (
+                "---\nExample:\n" f"## INPUT:\n{input_obj_text}\n" f"## OUTPUT:\n{self.object_to_text(output_obj)}\n"
+            )
+            prompt_clauses.append(prompt_clause)
+
+        prompt_end = "---\nQuery:\n" f"## INPUT:\n{query_text}\n" "## OUTPUT:\n"
+        system_prompt = SYSTEM_PROMPT.format(llm_config=self.config.llm_config)
+
+        def make_text(texts):
+            return "\n".join(prompt_clauses) + prompt_end
+
+        try:
+            encoding = encoding_for_model(model_name)
+        except KeyError:
+            encoding = encoding_for_model("gpt-4")
+        token_limit = get_token_limit(model_name)
+        prompt = render_formatted_text(make_text, prompt_clauses, encoding, token_limit)
+        logger.info(f"Prompt: {prompt}")
+        response = model.prompt(prompt, system_prompt)
+        yaml_str = response.text()
+        logger.info(f"Response: {yaml_str}")
+        return Inference(predicted_object=self._parse_yaml_payload(yaml_str))
+
+    def _parse_yaml_payload(self, yaml_str: str, strict=False) -> Optional[OBJECT]:
+        if "```" in yaml_str:
+            yaml_str = yaml_str.split("```")[1].strip()
+            if yaml_str.startswith("yaml"):
+                yaml_str = yaml_str[4:].strip()
+        try:
+            return yaml.safe_load(yaml_str)
+        except Exception as e:
+            if strict:
+                raise e
+            logger.error(f"Error parsing YAML: {yaml_str}\n{e}")
+            return None
+
+    def export_model(
+        self, output: Optional[Union[str, Path, TextIO]], model_serialization: ModelSerialization = None, **kwargs
+    ):
+        self.save_model(output)
+
+    def save_model(self, output: Union[str, Path]) -> None:
+        """
+        Save the trained model and related data to a file.
+
+        :param output: Path to save the model
+        """
+
+        # trigger index
+        _qr = self.rag_collection.search("*", limit=1)
+        assert len(_qr.ranked_rows) > 0
+
+        rows = self.rag_collection.find(limit=-1).rows
+
+        indexers = self.rag_collection.indexers
+        assert len(indexers) == 1
+        ix = self.rag_collection.indexers["llm"]
+        ix_coll = self.rag_collection.parent.get_collection(self.rag_collection.get_index_collection_name(ix))
+
+        ix_rows = ix_coll.find(limit=-1).rows
+        assert len(ix_rows) > 0
+        tm = TrainedModel(rag_collection_rows=rows, index_rows=ix_rows, config=self.config)
+        # tm = TrainedModel(rag_collection_rows=rows, index_rows=ix_rows)
+        with open(output, "w", encoding="utf-8") as f:
+            json.dump(tm.model_dump(), f)
+
+    @classmethod
+    def load_model(cls, file_path: Union[str, Path]) -> "RAGInferenceEngine":
+        """
+        Load a trained model and related data from a file.
+
+        :param file_path: Path to the saved model
+        :return: SklearnInferenceEngine instance with loaded model
+        """
+        with open(file_path, "r", encoding="utf-8") as f:
+            model_data = json.load(f)
+        tm = TrainedModel(**model_data)
+        from linkml_store.api import Client
+
+        client = Client()
+        db = client.attach_database("duckdb", alias="training")
+        db.store({"data": tm.rag_collection_rows})
+        collection = db.get_collection("data")
+        ix = collection.attach_indexer("llm", auto_index=False)
+        assert ix.name
+        ix_coll_name = collection.get_index_collection_name(ix)
+        assert ix_coll_name
+        ix_coll = db.get_collection(ix_coll_name, create_if_not_exists=True)
+        ix_coll.insert(tm.index_rows)
+        ie = cls(config=tm.config)
+        ie.rag_collection = collection
+        return ie

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/inference/implementations/sklearn_inference_engine.py

@@ -153,7 +153,7 @@ class SklearnInferenceEngine(InferenceEngine):
             y = y_encoder.fit_transform(y.values.ravel())  # Convert to 1D numpy array
             self.transformed_targets = y_encoder.classes_
 
-        logger.info(f"Fitting model with features: {X.columns}")
+        # print(f"Fitting model with features: {X.columns}")
         clf = DecisionTreeClassifier(random_state=42)
         clf.fit(X, y)
         self.classifier = clf

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/inference/inference_config.py

@@ -35,6 +35,7 @@ class InferenceConfig(BaseModel, extra="forbid"):
     feature_attributes: Optional[List[str]] = None
     train_test_split: Optional[Tuple[float, float]] = None
     llm_config: Optional[LLMConfig] = None
+    random_seed: Optional[int] = None
 
     @classmethod
     def from_file(cls, file_path: str, format: Optional[Format] = None) -> "InferenceConfig":

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/inference/inference_engine.py

@@ -29,6 +29,7 @@ class ModelSerialization(str, Enum):
     PNG = "png"
     LINKML_EXPRESSION = "linkml_expression"
     RULE_BASED = "rulebased"
+    RAG_INDEX = "rag_index"
 
     @classmethod
     def from_filepath(cls, file_path: str) -> Optional["ModelSerialization"]:
@@ -58,7 +59,7 @@ class ModelSerialization(str, Enum):
 
 
 class CollectionSlice(BaseModel):
-    model_config = ConfigDict(arbitrary_types_allowed=True)
+    model_config = ConfigDict(arbitrary_types_allowed=True, extra="forbid")
 
     name: Optional[str] = None
     base_collection: Optional[Collection] = None
@@ -69,17 +70,26 @@ class CollectionSlice(BaseModel):
 
     @property
     def collection(self) -> Collection:
+        if not self._collection and not self.indices:
+            return self.base_collection
         if not self._collection:
             rows = self.base_collection.find({}, limit=-1).rows
-            # subset based on indices
             subset = [rows[i] for i in self.indices]
             db = self.base_collection.parent
-            subset_name = f"{self.base_collection.alias}__rag_{self.name}"
+            subset_name = self.slice_alias
             subset_collection = db.get_collection(subset_name, create_if_not_exists=True)
+            # ensure the collection has the same schema type as the base collection;
+            # this ensures that column/attribute types are preserved
+            subset_collection.metadata.type = self.base_collection.target_class_name
+            subset_collection.delete_where({})
             subset_collection.insert(subset)
             self._collection = subset_collection
         return self._collection
 
+    @property
+    def slice_alias(self) -> str:
+        return f"{self.base_collection.alias}__rag_{self.name}"
+
     def as_dataframe(self, flattened=False) -> pd.DataFrame:
         """
         Return the slice of the collection as a dataframe.
@@ -113,31 +123,28 @@ class InferenceEngine(ABC):
 
         :param collection:
         :param split:
+        :param randomize:
         :return:
         """
+        local_random = random.Random(self.config.random_seed) if self.config.random_seed else random.Random()
         split = split or self.config.train_test_split
         if not split:
             split = (0.7, 0.3)
+        if split[0] == 1.0:
+            self.training_data = CollectionSlice(name="train", base_collection=collection, indices=None)
+            self.testing_data = None
+            return
         logger.info(f"Loading and splitting data from collection {collection.alias}")
         size = collection.size()
         indices = range(size)
         if randomize:
-            train_indices = random.sample(indices, int(size * split[0]))
+            train_indices = local_random.sample(indices, int(size * split[0]))
             test_indices = set(indices) - set(train_indices)
         else:
             train_indices = indices[: int(size * split[0])]
             test_indices = indices[int(size * split[0]) :]
         self.training_data = CollectionSlice(name="train", base_collection=collection, indices=train_indices)
         self.testing_data = CollectionSlice(name="test", base_collection=collection, indices=test_indices)
-        # all_data = collection.find({}, limit=size).rows
-        # all_data_df = nested_objects_to_dataframe(all_data)
-        # all_data_df = collection.find({}, limit=size).rows_dataframe
-        # randomize/shuffle order of rows in dataframe
-        # all_data_df = all_data_df.sample(frac=1).reset_index(drop=True)
-        # self.training_data = CollectionSlice(dataframe=all_data_df[: int(size * split[0])])
-        # self.testing_data = CollectionSlice(dataframe=all_data_df[int(size * split[0]) : size])
-        # self.training_data = CollectionSlice(base_collection=collection, slice=(0, int(size * split[0])))
-        # self.testing_data = CollectionSlice(base_collection=collection, slice=(int(size * split[0]), size))
 
     def initialize_model(self, **kwargs):
         """

{linkml_store-0.2.0 → linkml_store-0.2.1}/src/linkml_store/utils/llm_utils.py

@@ -20,6 +20,7 @@ MODEL_TOKEN_MAPPING = {
     "gpt-3.5-turbo-instruct": 4096,
     "text-ada-001": 2049,
     "ada": 2049,
+    "ada-002": 8192,
     "text-babbage-001": 2040,
     "babbage": 2049,
     "text-curie-001": 2049,

linkml_store-0.2.0/src/linkml_store/inference/implementations/rag_inference_engine.py

@@ -1,145 +0,0 @@
-import logging
-from dataclasses import dataclass
-from typing import Any, Optional
-
-import yaml
-from llm import get_key
-
-from linkml_store.api.collection import OBJECT, Collection
-from linkml_store.inference.inference_config import Inference, InferenceConfig, LLMConfig
-from linkml_store.inference.inference_engine import InferenceEngine
-from linkml_store.utils.object_utils import select_nested
-
-logger = logging.getLogger(__name__)
-
-SYSTEM_PROMPT = """
-You are a {llm_config.role}, your task is to inference the YAML
-object output given the YAML object input. I will provide you
-with a collection of examples that will provide guidance both
-on the desired structure of the response, as well as the kind
-of content.
-
-You should return ONLY valid YAML in your response.
-"""
-
-
-# def select_object(obj: OBJECT, key_paths: List[str]) -> OBJECT:
-# return {k: obj.get(k, None) for k in keys}
-# return {k: object_path_get(obj, k, None) for k in key_paths}
-
-
-@dataclass
-class RAGInferenceEngine(InferenceEngine):
-    """
-    AI Retrieval Augmented Generation (RAG) based predictor.
-
-
-    >>> from linkml_store.api.client import Client
-    >>> from linkml_store.utils.format_utils import Format
-    >>> from linkml_store.inference.inference_config import LLMConfig
-    >>> client = Client()
-    >>> db = client.attach_database("duckdb", alias="test")
-    >>> db.import_database("tests/input/countries/countries.jsonl", Format.JSONL, collection_name="countries")
-    >>> db.list_collection_names()
-    ['countries']
-    >>> collection = db.get_collection("countries")
-    >>> features = ["name"]
-    >>> targets = ["code", "capital", "continent", "languages"]
-    >>> llm_config = LLMConfig(model_name="gpt-4o-mini",)
-    >>> config = InferenceConfig(target_attributes=targets, feature_attributes=features, llm_config=llm_config)
-    >>> ie = RAGInferenceEngine(config=config)
-    >>> ie.load_and_split_data(collection)
-    >>> ie.initialize_model()
-    >>> prediction = ie.derive({"name": "Uruguay"})
-    >>> prediction.predicted_object
-    {'capital': 'Montevideo', 'code': 'UY', 'continent': 'South America', 'languages': ['Spanish']}
-
-    """
-
-    classifier: Any = None
-    encoders: dict = None
-    _model: "llm.Model" = None  # noqa: F821
-
-    rag_collection: Collection = None
-
-    def __post_init__(self):
-        if not self.config:
-            self.config = InferenceConfig()
-        if not self.config.llm_config:
-            self.config.llm_config = LLMConfig()
-
-    @property
-    def model(self) -> "llm.Model":  # noqa: F821
-        import llm
-
-        if self._model is None:
-            self._model = llm.get_model(self.config.llm_config.model_name)
-            if self._model.needs_key:
-                key = get_key(None, key_alias=self._model.needs_key)
-                self._model.key = key
-
-        return self._model
-
-    def initialize_model(self, **kwargs):
-        rag_collection = self.training_data.collection
-        rag_collection.attach_indexer("llm", auto_index=False)
-        self.rag_collection = rag_collection
-
-    def object_to_text(self, object: OBJECT) -> str:
-        return yaml.dump(object)
-
-    def derive(self, object: OBJECT) -> Optional[Inference]:
-        import llm
-        from tiktoken import encoding_for_model
-
-        from linkml_store.utils.llm_utils import get_token_limit, render_formatted_text
-
-        model: llm.Model = self.model
-        model_name = self.config.llm_config.model_name
-        feature_attributes = self.config.feature_attributes
-        target_attributes = self.config.target_attributes
-        num_examples = self.config.llm_config.number_of_few_shot_examples or 5
-        query_text = self.object_to_text(object)
-        if not self.rag_collection.indexers:
-            raise ValueError("RAG collection must have an indexer attached")
-        rs = self.rag_collection.search(query_text, limit=num_examples, index_name="llm")
-        examples = rs.rows
-        if not examples:
-            raise ValueError(f"No examples found for {query_text}; size = {self.rag_collection.size()}")
-        prompt_clauses = []
-        for example in examples:
-            # input_obj = {k: example.get(k, None) for k in feature_attributes}
-            input_obj = select_nested(example, feature_attributes)
-            # output_obj = {k: example.get(k, None) for k in target_attributes}
-            output_obj = select_nested(example, target_attributes)
-            prompt_clause = (
-                "---\nExample:\n"
-                f"## INPUT:\n{self.object_to_text(input_obj)}\n"
-                f"## OUTPUT:\n{self.object_to_text(output_obj)}\n"
-            )
-            prompt_clauses.append(prompt_clause)
-        # query_obj = {k: object.get(k, None) for k in feature_attributes}
-        query_obj = select_nested(object, feature_attributes)
-        query_text = self.object_to_text(query_obj)
-        prompt_end = "---\nQuery:\n" f"## INPUT:\n{query_text}\n" "## OUTPUT:\n"
-        system_prompt = SYSTEM_PROMPT.format(llm_config=self.config.llm_config)
-
-        def make_text(texts):
-            return "\n".join(prompt_clauses) + prompt_end
-
-        try:
-            encoding = encoding_for_model(model_name)
-        except KeyError:
-            encoding = encoding_for_model("gpt-4")
-        token_limit = get_token_limit(model_name)
-        prompt = render_formatted_text(make_text, prompt_clauses, encoding, token_limit)
-        logger.info(f"Prompt: {prompt}")
-        response = model.prompt(prompt, system_prompt)
-        yaml_str = response.text()
-        logger.info(f"Response: {yaml_str}")
-        try:
-            predicted_object = yaml.safe_load(yaml_str)
-            return Inference(predicted_object=predicted_object)
-        except yaml.parser.ParserError as e:
-            logger.error(f"Error parsing response: {yaml_str}\n{e}")
-            return None