§ 5Context Window Visualizer

Every LLM operates within a finite context window, a fixed budget of tokens that must accommodate system instructions, conversation history, retrieved knowledge, and space for the model's reply. When the window fills up, something must be evicted. This module visualizes the filling process in real time and compares four eviction strategies: FIFO, importance-based, recency-weighted, and LRU. Drag and drop segments to explore how prioritization changes what the model remembers and what it forgets.

Strategy

Evicts the oldest non-pinned segment first, regardless of its importance.

Context Window Size

Reserved for System Prompt200 tokens

0 tokens2000 tokens

Reserved for Response500 tokens

0 tokens2000 tokens

Available for Conversation3,396 tokens

Turn 0 / 30Auto-step

Figure 8

0 / 3,396 tokens used0.0% full

Turn 0 / 30

0 segments in window

system

user

assistant

retrieved

summary

Reserved

Real-time context window utilization. Each colored block represents a conversation segment proportional to its token count. Hatched zones are reserved for the system prompt and model response.

Figure 9

Drag to set priority (top = most important to keep)0 / 3,396 tokens

Drag segments to re-prioritize what stays in the context window. Pin critical segments to protect them from eviction. Lower items are evicted first when the window overflows.

Figure 10

FIFO

Best

Retained

Avg Importance

0.00

1st Eviction

Turn 0

Importance

Retained

Avg Importance

0.00

1st Eviction

Turn 0

Recency-Weighted

Retained

Avg Importance

0.00

1st Eviction

Turn 0

LRU

Retained

Avg Importance

0.00

1st Eviction

Turn 0

All four strategies process the same 0 segments with a 3,396-token available window. FIFO retains the highest average importance for this conversation.

Side-by-side comparison of all four eviction strategies applied to the same conversation. The strategy with the highest average importance retained is marked as best.

Key Insights

FIFO is simple but naive -- it discards early context even if it is highly important (e.g., critical instructions from turn 1).
Importance-based preserves value but can retain very old segments that may no longer be relevant to the current conversation topic.
Recency-weighted balances both axes -- it naturally deprioritizes old segments unless they are exceptionally important. This mirrors how human memory works.
LRU is ideal for interactive use -- segments the user has recently referenced stay in context, keeping the model responsive to the current thread of conversation.
Pinning is essential in practice -- system prompts, safety instructions, and user-defined facts should be pinned to avoid eviction regardless of strategy.
Reserved zones create a budget -- the system prompt and response reservation reduce the available space, making eviction more aggressive with smaller windows.

§ 5.5Validate Live: What Gets Evicted From Your Context Window?

Provide a conversation and set a token budget. The LLM decides what to keep and what to evict, showing the real tradeoffs of context window management on your data.

Provider

Model

API Key

Conversation

System: You are a travel planning assistant specializing in India.
User: I want to plan a trip to Kerala for 5 days in January.
Assistant: Kerala in January is perfect, it's post-monsoon with pleasant weather around 25-30°C. Would you like to focus on backwaters, hill stations, or beaches?
User: Mostly backwaters and some wildlife. My budget is ₹50,000 total.
Assistant: Great! I'd suggest: Days 1-2 in Alleppey for houseboats (₹8000/night), Day 3 at Periyar Wildlife Sanctuary in Thekkady, Days 4-5 in Munnar for tea plantations. Budget breakdown: houseboat ₹16000, Periyar ₹5000, Munnar hotel ₹8000, food ₹10000, transport ₹8000 = ₹47000.
User: My friend Deepa is joining. She has mobility issues, are houseboats accessible?
Assistant: Most Kerala houseboats are single-level and quite accessible. I'd recommend the premium houseboats which have wider walkways. For Periyar, skip the jungle trek and opt for the boat safari instead, much easier. Munnar has some steep terrain but many viewpoints are car-accessible.
User: Perfect. Also, Deepa is vegan, will that be an issue?
Assistant: Kerala is actually excellent for vegans! Traditional Kerala cuisine has many vegan dishes: avial, thoran, olan, and appam with coconut stew. Most houseboat kitchens can accommodate vegan requests easily.
User: What about the current weather forecast and any festivals happening?

~271 tokens estimated

Token budget: 200 tokens(71 tokens must be evicted)

50 tokens (tight)500 tokens (full)

§ 5Context Window Visualizer

Eviction Strategy

Window Configuration

Simulation Controls

§ 5.5Validate Live: What Gets Evicted From Your Context Window?