ContextGem

ContextGem splits the job in two. Aspects find whole sections, like termination clauses or payment terms, while Concepts pull the specific data points inside them, such as a monthly fee or a due date. That two-stage design keeps extraction focused: you narrow to the right part of a document before you ask for fine details, which makes results more accurate on long files and keeps the structured data clean.

The library ships with seven concept types: string, boolean, numerical, date, rating, JSON object, and label. Each one maps to a kind of answer, so a rating concept expects a number on a scale and a label concept picks from a list you set. Using typed concepts means the output arrives in the shape you need instead of a block of text you parse by hand.

Every extracted value keeps a link to the source text. Items carry paragraph references by default, and you can switch reference_depth to sentences for finer tracking. When a figure looks wrong during review, you can jump straight to the lines behind it rather than searching the whole document. Justifications are available too, so the model explains why it picked each item.

An LLM extraction pipeline bundles aspects and concepts into one reusable, serializable setup. You define it once, save it, and apply it to new documents with the same structure. Concurrency, retries, and cross-provider usage tracking are built in, so a batch of files runs without you managing the loop yourself. Serialization means a pipeline can move between machines or get stored as a versioned artifact.

ContextGem routes to any model that LiteLLM supports, covering OpenAI, Anthropic, Google, Azure, and xAI, plus local models through Ollama and LM Studio. That range lets you pick a fast, cheap model for simple pulls and a reasoning model for thorny clauses. Running locally keeps sensitive documents on your own hardware, which matters for legal and medical work.

You can ask the model to explain each extraction, with brief, balanced, or detailed justifications, plus a cap on sentence length. You can also attach examples to calibrate the model on your document type before it runs. Both features target accuracy: the examples teach the pattern, and the justifications let you audit the output instead of trusting it blind.