Back to Library
Use Cases

Your Support Search Index Dies on December 22: How a Forced Vector-Database Exit Becomes a Knowledge-Graph Support Desk

Last updated: November 20, 2026

Key takeaways

  • Chroma Cloud stops accepting writes on December 7, 2026 and serves its last read on December 22 — the vendor's own wind-down timeline sets both dates, 17 and 32 days from this article's publication.
  • A vector index freeze degrades support answers before it kills them — between December 7 and December 22 the desk still searches, but every new manual, release note, and warranty policy fails to enter the index, so agents answer from a stale corpus.
  • Graph-backed retrieval earned its evidence in production — LinkedIn's GraphRAG deployment improved retrieval accuracy (MRR) by 77.6% and cut issue resolution time by 28.6% (arXiv:2404.17723), which is the performance band a migration should aim at, not just a re-hosted vector index.
  • Open-source ChromaDB continues under Apache 2.0 — the wind-down applies only to the hosted cloud, and Chroma's acquisition announcement commits the team and Replit as contributors at roughly 6 million monthly downloads.

In 17 days, a hosted search index somewhere in your support stack stops accepting new information. Chroma — the embedded database behind a large share of the vector search systems built since 2023, at one point carrying what its CEO estimated as over 95% of vector-search downloads — signed a definitive agreement to join Replit on October 23, 2026, and the acquired service's wind-down timeline is unambiguous: writes disabled December 7, reads disabled December 22. The product survives as open-source ChromaDB; the hosted cloud your team may have picked for convenience two years ago does not.

Take a mid-market industrial-controls distributor: about 450 employees, roughly $200 million in revenue, running NetSuite for ERP and HubSpot Service Hub for tickets, with 9 support agents handling around 700 tickets a week against a knowledge base of some 12,000 documents — manuals, wiring diagrams, firmware release notes, warranty terms, all embedded into about 2.1 million vectors on Chroma Cloud. This article is for that team: what each deadline actually breaks, why the obvious fix (re-hosting the same vectors somewhere else) reproduces the problem you already had, and how the forced exit becomes the opening that gives the desk a knowledge graph it would never have budgeted for voluntarily.

What actually stops, and when

The wind-down page divides the exit into two boundaries, and they fail in different ways.

December 7 — writes disabled. From this date the index is read-only: no new tickets can be embedded, no new documents indexed, no deletions processed. For a support desk this is the more damaging line, and it arrives first. Any operational content with a short shelf life — a November firmware release note, a revised return policy, a supplier's end-of-life notice — simply never enters the index. Search keeps returning answers, which is exactly the trap: the desk sees no outage, only a slow drift toward stale answers that nobody can name until a customer holds the wrong revision of a datasheet.

December 22 — reads disabled. Sixty days after the October 23 notice, the endpoint stops answering entirely. After this date the application is not degraded; it is broken. If the embedding calls sit directly in the support widget path, the widget fails. If agents use search to resolve tickets, every ticket from December 23 onward is resolved the old way — by memory and folder-diving — until the migration lands.

Neither boundary is a surprise: the timeline was published October 23, and the two-week read-only window between December 7 and December 22 exists precisely so customers can export and validate data while the old answers still work. The acquisition announcement also closes off the "wait and see" path — Replit folds Chroma's infrastructure into its agent platform, so the hosted service your desk depends on is a strategic asset of a different product now, not a maintained utility.

The two-week freeze nobody plans for

Between the two deadlines sits a window that looks survivable and is not. The index still answers on December 10, so tickets resolve. But the corpus underneath is frozen at December 7, and support content is not. A distributor's December is its densest documentation season — year-end firmware refreshes, January pricing updates, revised warranty terms, corrected installation guides. Every one of these is invisible to retrieval after December 7.

The failure shows up as quality, not uptime: an agent quotes the superseded return window because the revision never got embedded; a suggested article macro recommends the pre-recall wiring sequence. The desk's escalation rate climbs a few points, nobody files an incident, and by January the team is debugging "the bot got worse" with a frozen index two layers below the symptom. The operational read is simple: a retrieval dependency with a dated end-of-life on it is a procurement event, not an engineering curiosity — and the team that owns the ticket queue needs to know before the write-stop, not after.

There is also a credentials step on the wind-down checklist worth naming: after December 22, rotate and remove the Chroma Cloud keys stored in your applications and infrastructure. An orphaned read credential pointed at a dead service is exactly the kind of thing a security review finds a year later.

The upgrade hiding in the deadline

The defensive move is mechanical: export the collections, re-host the vectors on another managed service or on self-hosted ChromaDB, repoint the client, ship before December 22. It works. It also reproduces the desk you already had, with a different vendor name on the same failure mode — an index that matches word overlap and cannot reason about how your products relate to each other.

The questions that actually cost a support desk its time are not similarity questions. They look like this:

  • "Which control board revision fits a 2019 furnace with the low-voltage accessory kit?" — a compatibility walk across product years, revisions, and option codes that no embedding distance resolves.
  • "Is this part still warrantied if the installer was third-party?" — entitlement logic that lives in policy text and business rules, not in prose similarity.
  • "What did the last four tickets on this asset class conclude?" — ticket history as structured memory, not as four more similar-sounding documents.

A knowledge graph encodes those relationships directly: products, revisions, compatibility edges, warranty rules, and resolved-ticket chains as nodes and typed edges, with HubSpot tickets and NetSuite part data connected as governed tools the retrieval agent calls — not scraped into text chunks. The agent that answers a ticket then has a path, not a guess: it walks the compatibility edge, checks the entitlement rule, cites the resolved-ticket chain, and returns the answer with its sources attached. Gated by no-citation-no-send, the desk's answers stay inspectable, which matters when the same graph starts feeding customer-facing self-service.

This is why the evidence for the graph-backed pattern is production-grade rather than benchmark-grade. The LinkedIn team's published system for customer-service question answering — a knowledge graph built over issue resolution chains and API dependencies with graph-augmented retrieval — improved retrieval accuracy (MRR) by 77.6% over the vector baseline and cut issue resolution time by 28.6% in production (arXiv:2404.17723). Those are the gains on the table when a desk migrates to a graph; a lift-and-shift migration offers only the floor of "answers keep coming."

The pragmatic sequence for this team fits inside the read-only window the vendor already granted:

  1. Before December 7 — redirect writes. Point new-ticket embedding and any document-sync job at the new home. Nothing new enters the old index.
  2. December 7–22 — export and stage. Export collections, migrate documents with their metadata, build the graph from the same corpus, and run graph-backed retrieval in shadow mode against live tickets while the frozen index remains the system of record.
  3. Cutover on your schedule, before December 22. Flip the widget and agent path to the graph-backed desk, compare shadow-mode answers against the frozen index, and rotate the Chroma Cloud credentials out of the application and infrastructure.

The outcome at support-desk scale

For a 9-agent desk the difference is measured in first-response and escalation, not in cloud credits. The answers the desk already struggles with — compatibility, entitlement, "what did we do last time" — move from a folder-diving escalation path (a human, 20–40 minutes, sometimes wrong) to a graph walk with citations (seconds, and inspectable). At the mid-market scale that is the hiring math that matters: the same 9 agents absorb growth or season peaks without a tenth hire, and the graph keeps absorbing new manuals and revised policies after December 7 — which, on the frozen timeline, nothing else could.

The vendor choice matters less than the pattern. Whether the graph lands in self-hosted Open-source ChromaDB's neighborhood alongside embeddings, in a dedicated graph store, or as a governed knowledge-graph module you own outright, the property to carry out of this exit is the same one the wind-down teaches: the retrieval layer is a dependency with a lifecycle, and the desk that owns its data model — not its vendor's SaaS — is the desk that migrates on its own calendar.

The Forced Exit Timeline — and What the Desk Does With It Chroma joins Replit (Oct 23, 2026) · writes stop Dec 7 · reads stop Dec 22 · open-source ChromaDB continues (Apache 2.0) OCT 23 Wind-down notice DEC 7 Writes disabled DEC 22 Reads disabled BEFORE — Frozen vector index Answers stop updating Dec 7 New manuals, policies, release notes: never indexed Word overlap, not reasoning "Which board fits option kit X?" → closest-sounding doc Offline entirely Dec 22 Widget and agent paths fail; folder-diving returns Quality drift, not an outage Escalation creep with no incident to file Lift-and-shift reproduction: same index, new vendor, same ceiling AFTER — Graph-backed support desk Answers keep absorbing Dec 7+ New manuals and revisions enter the graph on publish Compatibility + entitlement walks Products, revisions, warranty rules, ticket history as edges Cited answers, gated No-citation-no-send; audit path on every reply HubSpot + NetSuite as tools MCP modules the desk owns, not scraped text chunks Production evidence (arXiv:2404.17723): +77.6% retrieval accuracy · −28.6% resolution time The property to carry out of the exit: own the data model, not the vendor's SaaS. Sources: trychroma.com/cloud-wind-down · trychroma.com/blog/chroma-joins-replit · arXiv:2404.17723

The diagram renders the decision in one minute: two boundaries from the vendor's own timeline, the four ways a frozen vector index fails a support desk, and the graph-backed target state with the production evidence band.

Request a scoped build

Consider a representative build: the industrial-controls distributor above, 2.1 million frozen embeddings and a December 22 deadline. We export and validate the corpus inside the read-only window, stand up a knowledge graph over products, revisions, compatibility edges, warranty rules, and resolved-ticket chains; wrap HubSpot tickets and NetSuite part data as MCP modules with per-call audit logging; and run the graph-backed desk in shadow mode against live tickets before cutover — so the first day the old vendor is gone is the day the migration already finished.

Request a scoped build: one-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.

Related reading

Want this built for your systems?

Every document here comes from real production work. If you have a target system and a workflow in mind, we can scope a build in one week.

Request a scoped build

One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.