justinwhite.work

Context Sprawl Kills Momentum

In most organizations, product context is shattered across disconnected tools. Sherbet was built on the thesis that context is powerful; don't lose it.

by Justin White6 min read
Vol. I, No. OneReleased November 2026

The hardest part of shipping software rarely lies in writing the patch. It lies in reconstructing the context that explains why a bug happens and who it hurts.

In a typical engineering organization, that context is scattered across four or five silos. A user hits an edge case and drops a note into a feedback widget. Three more users email frontline support. A support agent asks a question in a chat channel. Finally, someone files a two-sentence ticket in an issue tracker, stripped of the original customer telemetry, the support conversation, and the fact that twenty other users reported the exact same symptom last week.

Context sprawl kills engineering momentum. Triage leads burn hours manually hunting for duplicates, engineers build fixes against incomplete reproduction steps, and customers never hear back when their issue is resolved.

Sherbet was built on a simple thesis: context is powerful; don't lose it. Instead of asking teams to manually tag, cross-reference, and triage every incoming signal, Sherbet treats context discovery as an intrinsic property of the platform.

Fingerprinting Every Primitive at Ingestion

In Sherbet, no piece of text sits in an isolated silo. Whenever any core entity is created or updated (whether a feedback report, support thread, issue, conversation, message, project, or document block), our ingestion pipeline runs it through

.

Each record is stamped with two lightweight, deterministic representations:

  • Compact Signatures: A serialized MinHash fingerprint capturing the normalized textual identity of the entity.
  • LSH Band Keys: An array of Locality-Sensitive Hashing band keys that group structurally similar problems into shared hash buckets, indexed directly in PostgreSQL using Generalized Inverted Indices (GIN).

Because generating a

Glyph
signature costs microseconds of CPU time and requires zero external GPU or embedding API calls, Sherbet fingerprints everything synchronously on write, even rolling up individual chat messages into live thread-level signatures as a support conversation unfolds.

Stamping signatures across every

primitive
enables a universal discovery engine exposed through Sherbet's query and related endpoints.

You can pass raw text, an existing entity ID, or a signature to hunt across the entire organization's knowledge graph. To keep multi-table discovery fast at scale, retrieval executes in two tiers:

  • Relational Candidate Pruning: PostgreSQL GIN indices evaluate array overlap across LSH band keys and tags in parallel across feedback, issues, support threads, and conversations. Instead of scanning thousands of historical rows, the database narrows the search space to a small candidate set sharing at least one band with the target.
  • In-Memory Reranking and Clustering: Candidate signatures are deserialized into an ephemeral in-memory Glyph index at the edge to compute exact Jaccard similarity scores, group matches into problem clusters, and calculate verified user impact by deduplicating customer emails across matching reports.

Context Hunting in Daily Workflow

Wiring related and query lookups directly into Sherbet's

workspace surfaces
changes how teams triage and build:

  • Automatic Duplicate Hunting: When an engineer opens an issue, Sherbet scans unlinked customer feedback and support threads automatically. A similarity drawer surfaces matching reports alongside their relevance scores, allowing the engineer to link all related signals in one click and see the true blast radius beside the ticket title.
  • Frontline Incident Elevation: When a support agent views an incoming customer thread, Sherbet immediately surfaces whether that problem is already tracked under an active engineering issue, preventing duplicate tickets and keeping the customer informed.
  • Proactive Ticket Deflection: As a user types into Sherbet's embeddable feedback widget, real-time queries match their input against known issues and solutions before a redundant ticket is even submitted.
  • Context Assembly for Coding Agents: Because related conversations, environment diagnostics, and customer reports are automatically clustered around each issue, AI coding agents can query a single endpoint to gather a complete reproduction spec before writing code.

When software discovers its own context, engineers stop acting as human routers between tools and get back to building.