Prompt Compression

The systematic removal of linguistically redundant tokens that do not contribute to the core legal instruction. This technique identifies and strips filler phrases, verbose transitions, and tautological statements that inflate prompt length without adding analytical value.

• Mechanism: Uses dependency parsing to isolate essential syntactic heads and discard modifiers that do not alter legal meaning
• Example: Compressing 'It is important to note that the contract clearly states that the party shall' to 'The contract states the party shall'
• Risk: Over-pruning can remove legally significant qualifiers like 'reasonable' or 'material'

A statistical compression method that evaluates each token's information-theoretic contribution to the prompt. Tokens with low self-information scores—those highly predictable from surrounding context—are candidates for removal.

• Mechanism: Calculates token-level entropy using the model's own probability distribution; tokens above a threshold are retained
• Application: Particularly effective for compressing boilerplate legal clauses where language is formulaic and highly predictable
• Trade-off: Requires an additional forward pass to compute token probabilities before compression

A two-stage pipeline where a smaller, faster language model first generates a condensed abstract of the source legal document, and this summary replaces the full text in the primary prompt. The technique trades minor fidelity loss for substantial token savings.

• Mechanism: A lightweight model (e.g., a fine-tuned T5 variant) produces a density-controlled summary before the main reasoning call
• Use Case: Compressing 50-page contracts into 2-page briefs for clause extraction tasks
• Critical Guardrail: The summarization model must be evaluated for citation fidelity to ensure no material obligations are dropped

The replacement of verbose instructional phrases with compact, high-density directives that the model has been demonstrated to interpret identically. This technique relies on a curated mapping of semantically equivalent prompt fragments.

• Mechanism: Maintains a lookup table mapping wordy instructions (e.g., 'You are required to carefully analyze and provide a detailed explanation of') to compact equivalents (e.g., 'Analyze and explain')
• Validation: Each synonym pair must be empirically tested via A/B Prompt Testing to confirm behavioral equivalence
• Benefit: Reduces instruction overhead by 30-50% without altering model behavior

A resource-allocation strategy that partitions the available context window into fixed token budgets for each prompt component: system instructions, retrieved documents, few-shot examples, and the user query. Compression is applied differentially based on each component's marginal utility.

• Mechanism: Assigns hard token limits per section; a dynamic allocator can shift budget from instructions to evidence when document length exceeds thresholds
• Legal Application: Ensures that in a Legal RAG Architecture, retrieved case law always receives sufficient token allocation even if it requires compressing the system prompt
• Implementation: Often integrated with Prompt Versioning systems to track budget configurations

A technique that replaces discrete natural language instructions with learned continuous vector embeddings optimized during fine-tuning. These 'soft prompts' occupy far fewer token slots while encoding complex behavioral directives.

• Mechanism: A small set of tunable embedding vectors (e.g., 5-20 tokens) is prepended to the input; these vectors are optimized via gradient descent to elicit target legal reasoning behavior
• Advantage: Compresses what would be a 500-token system prompt into 10 learned tokens
• Limitation: Soft prompts are model-specific and not human-readable, complicating Algorithmic Explainability audits