Next-gen RAG Observability

See exactly how your RAG pipeline is chunking your docs — before it breaks in production.

Compare chunking strategies side-by-side. Catch retrieval failures before your users do.

app.chunklens.dev/pipeline-eval?doc=api_spec_v2.pdf

[Product screenshot]

Strategy A: Semantic SplitterStrategy B: Recursive 512tStrategy C: Parent-Doc

Recall @ 5: 94.2% Latency: 42ms

Raw Document Stream (Token Heatmap)1,280 tokens

Chunk #01 · Tokens 0-340 · Topic: Auth & Sessions“OAuth2 token validation requires validating the asymmetric signature against the JWKS endpoint. Expired JWTs return HTTP 401 with code invalid_token...”

Chunk #02 · Tokens 341-690 · Topic: Rate Limiting Matrix“Tier 1 customers receive 10,000 req/min with token bucket bursts. Table: [Endpoint | Burst | Penalty]. Non-burst queries throttle at 200 req/s...”

Chunk #03 · Tokens 691-1020 · Topic: Error Fallbacks“When Redis is unavailable, the fallback circuit breaker switches to in-memory LRU cache with a 60s TTL...”

Evaluation Results+18.4% vs Baseline

Context Preservation96%

Table Boundary Integrity100%

Orphaned Chunk Bleed0% (Clean)

✓ 3 strategies benchmarked across 24 test queries.

ChunkLens identified 2 semantic break points preserved over fixed-size.

THE PROBLEM

Chunking is the least visible, most breakable part of RAG

Most teams treat document chunking as an afterthought — until hallucinations and poor top-k recall corrupt the user experience.

Flaky embeddings

Inconsistent rebuilds

The same document produces different chunks on every rebuild, and retrieval quality swings with it.

01 // failure_mode

Silent blindspots

No visibility

You can't see what your docs actually look like after chunking until something breaks downstream.

02 // failure_mode

Context fragmentation

Tables and structure break

Naive splitting divorces numbers from labels and breaks multi-section context.

03 // failure_mode

HOW IT WORKS

From raw docs to a scored comparison in minutes

No complex setup. Run your documents through multiple chunking engines and visualize the differences immediately.

Connect your docs

Upload a sample set or connect your existing document source.

Supports Markdown, PDF, HTML, JSON, and raw code repos.

Run every strategy at once

Fixed-size, semantic, and parent-document chunking, tested in parallel.

Evaluated simultaneously across custom overlap thresholds & embedding models.

See what actually changed

Visual diff of chunks plus a retrieval quality score against your own eval set.

Instant side-by-side visualizer with recall, latency, and context boundary metrics.

Built by an ML engineer who's shipped local LLM inference pipelines — for anyone building production RAG.

Join early engineers already on the list

Priority Beta Access

Get early access before public launch

Join engineers comparing chunking strategies, optimizing embeddings, and preventing silent retrieval failures.