Building LLMSlim: Architecture Deep-Dive into Deterministic Prompt Compression

Most prompt compression discussions focus on the happy path: you have a long RAG context, you trim it to 50% of tokens, and your API bill halves. What rarely gets discussed are the failure modes: dropped system instructions, truncated JSON schemas, broken code fences, and entity names that get quietly pruned because they scored low in some similarity metric.

LLMSlim (https://www.llmslim.app) is a Python library I've spent the last few months building to handle these edge cases properly. This post goes deep on the architecture decisions that make it work.

LLM input prompts have a structural problem. The instructions that govern model behavior (system roles, JSON schemas, MUST/NEVER directives) are typically a tiny fraction of the total token count. The rest is context: retrieved documents, conversation history, background information. And within that context, a significant fraction is prose that exists for human readability, not informational density.

The challenge is distinguishing high-value informational sentences from low-value filler without a model call, without embeddings, and in under 30ms.

LLMSlim processes prompts through a deterministic DAG (Directed Acyclic Graph) with six stages:

Naive sentence splitting on periods breaks code blocks and URLs. The first stage uses regex-based splitting that respects:

sentences = regex_split(text)
for i, sent in enumerate(sentences):
    if is_code_fence(sent): protected[i] = True  # immune to scoring

Each sentence becomes a TF-IDF vector. We compute the full pairwise cosine similarity matrix over these vectors to build a weighted undirected graph where edge weight(i,j) = cosine_similarity(v_i, v_j).

The choice of TF-IDF over neural embeddings was deliberate. TF-IDF runs in microseconds per sentence. It captures vocabulary overlap well, which is exactly what matters for identifying redundant prose. Neural embeddings add 50-100ms of latency and model overhead that isn't justified for this task.

We apply the LexRank algorithm: convert the similarity graph to a stochastic transition matrix M, then find the stationary distribution via power iteration:

p_t+1 = d * M * p_t + (1-d) / n

where d is a damping factor (typically 0.85) and n is the sentence count. Convergence typically happens in 20-30 iterations. The resulting stationary probability vector gives each sentence a centrality score representing how informationally central it is to the document.

This is the stage that separates LLMSlim from naive extraction approaches. Before any sentence can be pruned, a deterministic rule pass classifies every sentence into one of four tiers:

Tier 4 (Inviolable): Sentences containing:

system:

, developer:

, user:

MUST

, NEVER

, ALWAYS

, REQUIRED

, DO NOT

Tier 3 (Protected): Sentences containing:

Tier 2 (Standard): Regular content sentences

Tier 1 (Candidate for removal): Filler phrases, transition sentences

Tier 4 sentences are hardcoded to survive the compression pass regardless of their LexRank score. A sentence saying "You MUST respond only in JSON" will score low in a document full of prose paragraphs - but it's the most important sentence in the prompt.

Pass 1 divides the document into semantic chunks and allocates token budgets proportionally:

chunk_budget_i = target_tokens * (chunk_tokens_i / total_tokens)

Pass 2 applies a global rebalancing step. Each chunk may have selected sentences that, together, slightly exceed or undershoot its budget. The second pass collects the surplus/deficit across all chunks and redistributes tokens using a priority-aware greedy knapsack:

This two-pass approach is why the library consistently hits within 2-3% of the target ratio even on highly variable document structures.

Selected sentences are sorted by their original document position and concatenated. Original ordering is preserved because causal reasoning and logical flow in documents depends on sequence, not just content.

v0.3.0 adds a generative compression layer on top of the extractive pipeline. The flow:

The pluggable provider model means you don't need a specific LLM API key:

from llmslim import compress, CallableProvider, RewriteRequest

def my_provider(req: RewriteRequest) -> str:
    return your_llm_function(req.user_prompt)

provider = CallableProvider(my_provider)
result = compress(context, target_ratio=0.4, strategy="hybrid", provider=provider)

All benchmarks are reproducible. Hardware: AMD EPYC 7763, 64GB RAM, Ubuntu 24.04, Python 3.12.3, tiktoken cl100k_base. N=500 prompts per dataset, 100 iterations per sample.

Attempt 1: Pure TF-IDF cutoff. Setting a similarity threshold and dropping sentences below it sounds simple. In practice, document sections with specialized vocabulary score low across the board even when they contain critical information. Threshold tuning became document-specific.

Attempt 2: Summarization for compression. Running a smaller model to summarize chunks seems appealing. It adds a full model inference call (50-500ms), requires an API dependency, and summaries tend to drop entity specifics that matter for downstream model performance.

Attempt 3: Single-pass budget allocation. Works fine for uniform documents. Fails badly on documents with mixed density: technical specs mixed with background narrative. The two-pass approach emerged from debugging these failures.

pip install llmslim

The benchmark scripts, raw JSON payloads, and full methodology are open. If you find issues with the numbers, open an issue - I'd genuinely like to know.