Skip to main content
V2 is available on Pro, Team, and Enterprise plans. Upgrade your plan to get access.
V2 keeps everything you already know from V1 and adds a set of capabilities that become important as your workload grows: semantic retrieval that understands meaning rather than just keywords, serving profiles that let you trade latency for recall on a per-request basis, an adaptive budget that expands when evidence is sparse, and async ingest with job tracking so indexing never blocks your hot path.

​
V1 vs V2 at a glance

V1V2
RetrievalKeyword + temporalKeyword + temporal + semantic
Serving profiles—low_latency, balanced, high_recall
Token budgetFixedAdaptive — expands when needed
IngestSingle artifactBatch array in one call
Async indexing—Per-request async_index flag + job tracking
Workspaces1Pro: 5 · Team: 20 · Enterprise: unlimited
Token budget (max)4,09616,384 – 131,072
RPM30120 – 1,000

​
Serving profiles

V2 lets you pick a retrieval strategy per request instead of applying one strategy to everything.

low_latency

Fastest response. Skips semantic search. Best for real-time chat and quick lookups where a slightly smaller result set is fine.

balanced

The default. Semantic search on with adaptive budget. Covers most applications well.

high_recall

Maximum evidence. Runs multiple retrieval passes with an expanded budget. Best for compliance, legal review, and research.
Pass the profile on any /v2/context-pack request: 
pack = client.context_pack("What changed in Q3 pricing?", profile="high_recall")
The response tells you which profile was actually used (the system can escalate if it determines results are sparse) and whether any degradation occurred.

​
Semantic retrieval

V1 retrieval matches on keywords and recency. V2 adds a semantic channel that understands meaning — so a query like “customer complaints about billing” also surfaces artifacts that talk about “payment disputes” or “invoice errors” even if those exact words don’t appear in your query. Semantic retrieval is on by default in balanced and high_recall profiles. It’s off in low_latency to keep latency predictable.

​
Adaptive budget

In V1, your max_tokens is a hard ceiling. In V2, the runtime can expand that budget automatically when evidence is sparse — so you get a meaningful context pack even when your memory doesn’t have dense coverage of the query topic. The expansion multiplier depends on your serving profile:
ProfileAdaptive budget
low_latencyOff
balancedUp to 1.5×
high_recallUp to 4×, capped at 4,096 tokens
You always see the actual token count in the token_accounting field of the response — the budget never silently exceeds your account’s plan limit.

​
Batch ingest and async job tracking

V1 ingests one artifact per request. V2 accepts an array, and with async_index: true you get job IDs back immediately — no waiting for indexing to complete before moving on. 
# Ingest a batch of 20 chat turns in one call
result = client.ingest([
    {"artifact_type": "chat_turn", "raw_payload": {"role": "user", "content": msg}}
    for msg in conversation
], async_index=True)

# Poll until done
for job in result["queued_jobs"]:
    status = client.job_status(job["job_id"])
    print(f"{job['job_id']}: {status['status']}")
See POST /v2/ingest and GET /v2/jobs/ for the full reference.

​
Multiple workspaces

V1 gives every tenant one workspace. V2 lets you create named workspaces — each fully isolated, each routable to its own storage if you’re on an appropriate plan.
PlanWorkspaces
Pro5
Team20
EnterpriseUnlimited
Use workspaces to separate environments (staging vs. prod), projects, or tenants in a multi-tenant product.

Upgrade to V2

Pro starts at $9.99/mo. Everything in Free, plus V2 access, semantic retrieval, adaptive budget, and 100K events/month.