The spacy-llm package integrates Large Language Models (LLMs) into spaCy, featuring a modular system for fast prototyping and prompting, and turning unstructured responses into robust outputs for various NLP tasks, no training data required.

Config and implementation

An LLM component is implemented through the LLMWrapper class. It is accessible through a generic llm component factory as well as through task-specific component factories: llm_ner, llm_spancat, llm_rel, llm_textcat, llm_sentiment, llm_summarization, llm_entity_linker, llm_raw and llm_translation. For these factories, the GPT-3-5 model from OpenAI is used by default, but this can be customized.

LLMWrapper.__init__ method

Create a new pipeline instance. In your application, you would normally use a shortcut for this and instantiate the component using its string name and nlp.add_pipe.

LLMWrapper.__call__ method

Apply the pipe to one document. The document is modified in place and returned. This usually happens under the hood when the nlp object is called on a text and all pipeline components are applied to the Doc in order.

LLMWrapper.pipe method

Apply the pipe to a stream of documents. This usually happens under the hood when the nlp object is called on a text and all pipeline components are applied to the Doc in order.

LLMWrapper.add_label method

Add a new label to the pipe’s task. Alternatively, provide the labels upon the task definition, or through the [initialize] block of the config.

LLMWrapper.to_disk method

Serialize the pipe to disk.

LLMWrapper.from_disk method

Load the pipe from disk. Modifies the object in place and returns it.

LLMWrapper.to_bytes method

Serialize the pipe to a bytestring.

LLMWrapper.from_bytes method

Load the pipe from a bytestring. Modifies the object in place and returns it.

LLMWrapper.labels property

The labels currently added to the component. Empty tuple if the LLM’s task does not require labels.

Tasks

In spacy-llm, a task defines an NLP problem or question and its solution using an LLM. It does so by implementing the following responsibilities:

1. Loading a prompt template and injecting documents’ data into the prompt. Optionally, include fewshot examples in the prompt.
2. Splitting the prompt into several pieces following a map-reduce paradigm, if the prompt is too long to fit into the model’s context and the task supports sharding prompts.
3. Parsing the LLM’s responses back into structured information and validating the parsed output.

Two different task interfaces are supported: ShardingLLMTask and NonShardingLLMTask. Only the former supports the sharding of documents, i. e. splitting up prompts if they are too long.

All tasks are registered in the llm_tasks registry.

On Sharding

“Sharding” describes, generally speaking, the process of distributing parts of a dataset across multiple storage units for easier processing and lookups. In spacy-llm we use this term (synonymously: “mapping”) to describe the splitting up of prompts if they are too long for a model to handle, and “fusing” (synonymously: “reducing”) to describe how the model responses for several shards are merged back together into a single document.

Prompts are broken up in a manner that always keeps the prompt in the template intact, meaning that the instructions to the LLM will always stay complete. The document content however will be split, if the length of the fully rendered prompt exceeds a model context length.

A toy example: let’s assume a model has a context window of 25 tokens and the prompt template for our fictional, sharding-supporting task looks like this:

Depending on how tokens are counted exactly (this is a config setting), we might come up with n = 12 tokens for the number of tokens in the prompt instructions. Furthermore let’s assume that our text is “This has been amazing - I can’t remember the last time I left the cinema so impressed.” - which has roughly 19 tokens.

Considering we only have 13 tokens to add to our prompt before we hit the context limit, we’ll have to split our prompt into two parts. Thus spacy-llm, assuming the task used supports sharding, will split the prompt into two (the default splitting strategy splits by tokens, but alternative splitting strategies splitting e. g. by sentences can be configured):

(Prompt 1/2)

(Prompt 2/2)

The reduction step is task-specific - a sentiment estimation task might e. g. do a weighted average of the sentiment scores. Note that prompt sharding introduces potential inaccuracies, as the LLM won’t have access to the entire document at once. Depending on your use case this might or might not be problematic.

NonShardingLLMTask

task.generate_prompts

Takes a collection of documents, and returns a collection of “prompts”, which can be of type Any. Often, prompts are of type str - but this is not enforced to allow for maximum flexibility in the framework.

task.parse_responses

Takes a collection of LLM responses and the original documents, parses the responses into structured information, and sets the annotations on the documents. The parse_responses function is free to set the annotations in any way, including Doc fields like ents, spans or cats, or using custom defined fields.

The responses are of type Iterable[Any], though they will often be str objects. This depends on the return type of the model.

ShardingLLMTask

task.generate_prompts

Takes a collection of documents, breaks them up into shards if necessary to fit all content into the model’s context, and returns a collection of collections of “prompts” (i. e. each doc can have multiple shards, each of which have exactly one prompt), which can be of type Any. Often, prompts are of type str - but this is not enforced to allow for maximum flexibility in the framework.

task.parse_responses

Receives a collection of collections of LLM responses (i. e. each doc can have multiple shards, each of which have exactly one prompt / prompt response) and the original shards, parses the responses into structured information, sets the annotations on the shards, and merges back doc shards into single docs. The parse_responses function is free to set the annotations in any way, including Doc fields like ents, spans or cats, or using custom defined fields.

The responses are of type Iterable[Iterable[Any]], though they will often be str objects. This depends on the return type of the model.

Translation

The translation task translates texts from a defined or inferred source to a defined target language.

spacy.Translation.v1

spacy.Translation.v1 supports both zero-shot and few-shot prompting.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

Raw prompting

Different to all other tasks spacy.Raw.vX doesn’t provide a specific prompt, wrapping doc data, to the model. Instead it instructs the model to reply to the doc content. This is handy for use cases like question answering (where each doc contains one question) or if you want to include customized prompts for each doc.

spacy.Raw.v1

Note that since this task may request arbitrary information, it doesn’t do any parsing per se - the model response is stored in a custom Doc attribute (i. e. can be accessed via doc._.{field}).

It supports both zero-shot and few-shot prompting.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

Summarization

A summarization task takes a document as input and generates a summary that is stored in an extension attribute.

spacy.Summarization.v1

The spacy.Summarization.v1 task supports both zero-shot and few-shot prompting.

The summarization task prompts the model for a concise summary of the provided text. It optionally allows to limit the response to a certain number of tokens - note that this requirement will be included in the prompt, but the task doesn’t perform a hard cut-off. It’s hence possible that your summary exceeds max_n_words.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

EL (Entity Linking)

The EL links recognized entities (see NER) to those in a knowledge base (KB). The EL task prompts the LLM to select the most likely candidate from the KB, whose structure can be arbitrary.

Note that the documents processed by the entity linking task are expected to have recognized entities in their .ents attribute. This can be achieved by either running the NER task, using a trained spaCy NER model or setting the entities manually prior to running the EL task.

In order to be able to pull data from the KB, an object implementing the CandidateSelector protocol has to be provided. This requires two functions: (1) __call__() to fetch candidate entities for entity mentions in the text (assumed to be available in Doc.ents) and (2) get_entity_description() to fetch descriptions for any given entity ID. Descriptions can be empty, but ideally provide more context for entities stored in the KB.

spacy-llm provides a CandidateSelector implementation (spacy.CandidateSelector.v1) that leverages a spaCy knowledge base - as used in an entity_linking component - to select candidates. This knowledge base can be loaded from an existing spaCy pipeline (note that the pipeline’s EL component doesn’t have to be trained) or from a separate .yaml file.

spacy.EntityLinker.v1

Supports zero- and few-shot prompting. Relies on a configurable component suggesting viable entities before letting the LLM pick the most likely candidate.

spacy.CandidateSelector.v1

spacy.CandidateSelector.v1 is an implementation of the CandidateSelector protocol required by spacy.EntityLinker.v1. The built-in candidate selector method allows loading existing knowledge bases in several ways, e. g. loading from a spaCy pipeline with a (not necessarily trained) entity linking component, and loading from a file describing the knowlege base as a .yaml file. Either way the loaded data will be converted to a spaCy InMemoryLookupKB instance. The KB’s selection capabilities are used to select the most likely entity candidates for the specified mentions.

spacy.KBObjectLoader.v1

Adheres to the InMemoryLookupKBLoader interface required by spacy.CandidateSelector.v1. Loads a knowledge base from an existing spaCy pipeline.

spacy.KBFileLoader.v1

Adheres to the InMemoryLookupKBLoader interface required by spacy.CandidateSelector.v1. Loads a knowledge base from a knowledge base file. The KB .yaml file has to stick to the following format:

See here for a toy example of how such a KB file might look like.

NER

The NER task identifies non-overlapping entities in text.

spacy.NER.v3

Version 3 is fundamentally different to v1 and v2, as it implements Chain-of-Thought prompting, based on the PromptNER paper by Ashok and Lipton (2023). On an internal use-case, we have found this implementation to obtain significant better accuracy - with an increase of F-score of up to 15 percentage points.

When no examples are specified, the v3 implementation will use a dummy example in the prompt. Technically this means that the task will always perform few-shot prompting under the hood.

Note that the single_match parameter, used in v1 and v2, is not supported anymore, as the CoT parsing algorithm takes care of this automatically.

New to v3 is the fact that you can provide an explicit description of what entities should look like. You can use this feature in addition to label_definitions.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

While not required, this task works best when both positive and negative examples are provided. The format is different than the files required for v1 and v2, as additional fields such as is_entity and reason should now be provided.

For a fully working example, see this usage example.

spacy.NER.v2

This version supports explicitly defining the provided labels with custom descriptions, and further supports zero-shot and few-shot prompting just like v1.

The parameters alignment_mode, case_sensitive_matching and single_match are identical to the v1 implementation. The format of few-shot examples are also the same.

New to v2 is the fact that you can write definitions for each label and provide them via the label_definitions argument. This lets you tell the LLM exactly what you’re looking for rather than relying on the LLM to interpret its task given just the label name. Label descriptions are freeform so you can write whatever you want here, but a brief description along with some examples and counter examples seems to work quite well.

For a fully working example, see this usage example.

spacy.NER.v1

The original version of the built-in NER task supports both zero-shot and few-shot prompting.

The NER task implementation doesn’t currently ask the LLM for specific offsets, but simply expects a list of strings that represent the enties in the document. This means that a form of string matching is required. This can be configured by the following parameters:

• The single_match parameter is typically set to False to allow for multiple matches. For instance, the response from the LLM might only mention the entity “Paris” once, but you’d still want to mark it every time it occurs in the document.
• The case-sensitive matching is typically set to False to be robust against case variances in the LLM’s output.
• The alignment_mode argument is used to match entities as returned by the LLM to the tokens from the original Doc - specifically it’s used as argument in the call to doc.char_span(). The "strict" mode will only keep spans that strictly adhere to the given token boundaries. "contract" will only keep those tokens that are fully within the given range, e.g. reducing "New Y" to "New". Finally, "expand" will expand the span to the next token boundaries, e.g. expanding "New Y" out to "New York".

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

SpanCat

The SpanCat task identifies potentially overlapping entities in text.

spacy.SpanCat.v3

The built-in SpanCat v3 task is a simple adaptation of the NER v3 task to support overlapping entities and store its annotations in doc.spans.

Note that the single_match parameter, used in v1 and v2, is not supported anymore, as the CoT parsing algorithm takes care of this automatically.

spacy.SpanCat.v2

The built-in SpanCat v2 task is a simple adaptation of the NER v2 task to support overlapping entities and store its annotations in doc.spans.

Except for the spans_key parameter, the SpanCat v2 task reuses the configuration from the NER v2 task. Refer to its documentation for more insight.

spacy.SpanCat.v1

The original version of the built-in SpanCat task is a simple adaptation of the v1 NER task to support overlapping entities and store its annotations in doc.spans.

Except for the spans_key parameter, the SpanCat v1 task reuses the configuration from the NER v1 task. Refer to its documentation for more insight.

TextCat

The TextCat task labels documents with relevant categories.

spacy.TextCat.v3

On top of the functionality from v2, version 3 of the built-in TextCat tasks allows setting definitions of labels. Those definitions are included in the prompt.

The formatting of few-shot examples is the same as those for the v1 implementation.

spacy.TextCat.v2

V2 includes all v1 functionality, with an improved prompt template.

The formatting of few-shot examples is the same as those for the v1 implementation.

spacy.TextCat.v1

Version 1 of the built-in TextCat task supports both zero-shot and few-shot prompting.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

If you want to perform few-shot learning with a binary classifier (i. e. a text either should or should not be assigned to a given class), you can provide positive and negative examples with answers of “POS” or “NEG”. “POS” means that this example should be assigned the class label defined in the configuration, “NEG” means it shouldn’t. E. g. for spam classification:

REL

The REL task extracts relations between named entities.

spacy.REL.v1

The built-in REL task supports both zero-shot and few-shot prompting. It relies on an upstream NER component for entities extraction.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

Note: the REL task relies on pre-extracted entities to make its prediction. Hence, you’ll need to add a component that populates doc.ents with recognized spans to your spaCy pipeline and put it before the REL component.

For a fully working example, see this usage example.

Lemma

The Lemma task lemmatizes the provided text and updates the lemma_ attribute in the doc’s tokens accordingly.

spacy.Lemma.v1

This task supports both zero-shot and few-shot prompting.

The task prompts the LLM to lemmatize the passed text and return the lemmatized version as a list of tokens and their corresponding lemma. E. g. the text I'm buying ice cream for my friends should invoke the response

If for any given text/doc instance the number of lemmas returned by the LLM doesn’t match the number of tokens from the pipeline’s tokenizer, no lemmas are stored in the corresponding doc’s tokens. Otherwise the tokens .lemma_ property is updated with the lemma suggested by the LLM.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

Sentiment

Performs sentiment analysis on provided texts. Scores between 0 and 1 are stored in Doc._.sentiment - the higher, the more positive. Note in cases of parsing issues (e. g. in case of unexpected LLM responses) the value might be None.

spacy.Sentiment.v1

This task supports both zero-shot and few-shot prompting.

To perform few-shot learning, you can write down a few examples in a separate file, and provide these to be injected into the prompt to the LLM. The default reader spacy.FewShotReader.v1 supports .yml, .yaml, .json and .jsonl.

NoOp

This task is only useful for testing - it tells the LLM to do nothing, and does not set any fields on the docs.

spacy.NoOp.v1

This task needs no further configuration.

Models

A model defines which LLM model to query, and how to query it. It can be a simple function taking a collection of prompts (consistent with the output type of task.generate_prompts()) and returning a collection of responses (consistent with the expected input of parse_responses). Generally speaking, it’s a function of type Callable[[Iterable[Iterable[Any]]], Iterable[Iterable[Any]]], but specific implementations can have other signatures, like Callable[[Iterable[Iterable[str]]], Iterable[Iterable[str]]].

Note: the model signature expects a nested iterable so it’s able to deal with sharded docs. Unsharded docs (i. e. those produced by (nonsharding tasks)[/api/large-language-models#task-nonsharding]) are reshaped to fit the expected data structure.

Models via REST API

These models all take the same parameters, but note that the config should contain provider-specific keys and values, as it will be passed onwards to the provider’s API.

Currently, these models are provided as part of the core library:

To use these models, make sure that you’ve set the relevant API keys as environment variables.

⚠️ A note on spacy.Azure.v1. Working with Azure OpenAI is slightly different than working with models from other providers:

• In Azure LLMs have to be made available by creating a deployment of a given model (e. g. GPT-3.5). This deployment can have an arbitrary name. The name argument, which everywhere else denotes the model name (e. g. claude-1.0, gpt-3.5), here refers to the deployment name.
• Deployed Azure OpenAI models are reachable via a resource-specific base URL, usually of the form https://{resource}.openai.azure.com. Hence the URL has to be specified via the base_url argument.
• Azure further expects the API version to be specified. The default value for this, via the api_version argument, is currently 2023-05-15 but may be updated in the future.
• Finally, since we can’t infer information about the model from the deployment name, spacy-llm requires the model_type to be set to either "completions" or "chat", depending on whether the deployed model is a completion or chat model.

API Keys

Note that when using hosted services, you have to ensure that the proper API keys are set as environment variables as described by the corresponding provider’s documentation.

E. g. when using OpenAI, you have to get an API key from openai.com, and ensure that the keys are set as environmental variables:

For Cohere:

For Anthropic:

For PaLM:

Models via HuggingFace

These models all take the same parameters:

Currently, these models are provided as part of the core library:

Note that Hugging Face will download the model the first time you use it - you can define the cached directory by setting the environmental variable HF_HOME.

Installation with HuggingFace

To use models from HuggingFace, ideally you have a GPU enabled and have installed transformers, torch and CUDA in your virtual environment. This allows you to have the setting device=cuda:0 in your config, which ensures that the model is loaded entirely on the GPU (and fails otherwise).

You can do so with

If you don’t have access to a GPU, you can install accelerate and setdevice_map=auto instead, but be aware that this may result in some layers getting distributed to the CPU or even the hard drive, which may ultimately result in extremely slow queries.

LangChain models

To use LangChain for the API retrieval part, make sure you have installed it first:

Note that LangChain currently only supports Python 3.9 and beyond.

LangChain models in spacy-llm work slightly differently. langchain’s models are parsed automatically, each LLM class in langchain has one entry in spacy-llm’s registry. As langchain’s design has one class per API and not per model, this results in registry entries like langchain.OpenAI.v1 - i. e. there is one registry entry per API and not per model (family), as for the REST- and HuggingFace-based entries.

The name of the model to be used has to be passed in via the name attribute.

The default query (spacy.CallLangChain.v1) executes the prompts by running model(text) for each given textual prompt.

Cache

Interacting with LLMs, either through an external API or a local instance, is costly. Since developing an NLP pipeline generally means a lot of exploration and prototyping, spacy-llm implements a built-in cache to avoid reprocessing the same documents at each run that keeps batches of documents stored on disk.

When retrieving a document, the BatchCache will first figure out what batch the document belongs to. If the batch isn’t in memory it will try to load the batch from disk and then move it into memory.

Note that since the cache is generated by a registered function, you can also provide your own registered function returning your own cache implementation. If you wish to do so, ensure that your cache object adheres to the Protocol defined in spacy_llm.ty.Cache.

Various functions

spacy.FewShotReader.v1

This function is registered in spaCy’s misc registry, and reads in examples from a .yml, .yaml, .json or .jsonl file. It uses srsly to read in these files and parses them depending on the file extension.

spacy.FileReader.v1

This function is registered in spaCy’s misc registry, and reads a file provided to the path to return a str representation of its contents. This function is typically used to read Jinja files containing the prompt template.

Normalizer functions

These functions provide simple normalizations for string comparisons, e.g. between a list of specified labels and a label given in the raw text of the LLM response. They are registered in spaCy’s misc registry and have the signature Callable[[str], str].

• spacy.StripNormalizer.v1: only apply text.strip()
• spacy.LowercaseNormalizer.v1: applies text.strip().lower() to compare strings in a case-insensitive way.