How do I make a prompt fit in an 8K context window?

Paste the prompt here, select 8K as the target window, enter the question you're about to ask, and choose Smart strategy. We keep the system prompt verbatim, preserve the last 3 messages, keep whatever is relevant to your question, and drop or summarize the rest until the result fits 8,000 tokens with 20% held back for the answer. This is exactly what fixes a "prompt tokens exceed context window" or "this model's maximum context length is 8192 tokens" error.

What is the difference between Smart Token Reducer and Prompt Compressor?

Smart Token Reducer is lossless — it strips whitespace, filler words, typos, emoji, and casing waste, and the meaning never changes. It typically saves 20-40% and needs no question. Prompt Compressor is lossy but intelligent — it removes or summarizes whole paragraphs based on relevance to your question, saves 50-90%, and is built specifically to hit a target token count. Use the Reducer when you just want cheaper; use the Compressor when the prompt genuinely doesn't fit.

Will compression hurt the answer quality?

Some information is necessarily lost — that's what makes this lossy. Microsoft's LongLLMLingua research found that when compression is question-aware (biased toward content relevant to what you're about to ask), models can lose under 5% quality even at high compression ratios, because irrelevant padding was never helping the answer anyway. Smart mode targets that low-loss range; Aggressive mode trades more quality for a smaller prompt.

Can I make a 100K-token prompt fit into a 32K window?

Yes — set the target to 32K, pick Aggressive strategy, and enter your question. A 100K prompt typically compresses to roughly 25-30K (70-75% savings) in Smart mode, more in Aggressive mode. For very large sources like a full codebase or a multi-document knowledge base, chunking for RAG (see Docs to LLM) is usually the better long-term fix — compression is best for "I need this one prompt to fit right now".

Does it preserve my code blocks and tables?

Yes, by default. Code Blocks and Tables are on in Preserve Options — fenced code (```) and Markdown tables are never touched by the compressor, regardless of how the compression level or strategy is set. Turn a toggle off only if you specifically want that content eligible for removal.

How does question-aware compression actually work?

You type the question or task you're about to send with the prompt — e.g. "What is the refund window for damaged items?" We extract its keywords, score every paragraph in your prompt by how much it overlaps with those keywords, and remove or summarize the lowest-scoring paragraphs first. Paragraphs that mention refunds and damaged items survive; unrelated boilerplate goes first. Leave the question blank and we fall back to a positional heuristic — keep the start and end, cut the middle — since models tend to pay less attention to the middle of a long context anyway.

What does the Compression Level slider actually control?

It sets a floor, independent of whether the prompt already fits: at 30% (Conservative) we keep roughly 70% of the original; at 80% (Aggressive) we keep roughly 20%. The tool always uses whichever is stricter — the level's floor, or what the target window actually requires — so a prompt that already fits won't be touched unless the level asks for more.