Skip to main content
​
2026-07-16 · temporal annotation format fix

​
Fixed: temporal annotation format on PATCH /api/v1/tkgs/{tkgId}/graph/nodes/{nodeId}/annotation

The request body now uses { timestamp: "2026-09-30", precision: "quarter" } (an ISO 8601 date) instead of the previous loose { when: "2026 Q3", precision: "quarter" } string. This aligns with the annotation format already used elsewhere on the platform.The endpoint now also merges into any existing annotation data instead of overwriting it, so other fields (like source provenance) are preserved.capture_temporal actions from POST /api/v1/chat now return timestamp (ISO 8601) instead of when, so you can pass it straight through to this endpoint.
​
2026-07-16 · title dedup on node creation

​
Improved: duplicate-safe graph node creation

POST /api/v1/collections/{id}/graph/nodes and POST /api/v1/tkgs/{tkgId}/graph/nodes now check for an existing prediction/claim with a substantially overlapping title before creating a new one. If a close match already exists on the platform, that existing FactBlock is reused instead of creating a near-duplicate — and if it’s already a node in the same graph, no duplicate node is created either.The response now includes an optional reused: boolean field so you can tell whether your request created something new or matched an existing FactBlock.
​
2026-07-16 · TKG-grounded chat improvements

​
Improved: tkg_id grounding in POST /api/v1/chat

When chatting with tkg_id grounding, the assistant now works much harder to help you build out your knowledge graph:
  • New capture_temporal action — alongside capture_belief and capture_causal, the assistant can now suggest attaching a time reference (when + precision: year/quarter/month/day) to an existing belief.
  • Graph-gap awareness — the assistant is now aware of which beliefs have no causal connections and which predictions have no timeline, and will proactively ask about them.
  • FactBlock reuse — the assistant checks for related predictions/claims that already exist on the platform before treating something as a brand-new belief.
  • Implicit capture from free text — if you mention a specific time or a causal reason in your own message (not just by clicking a suggested action), the assistant will include the matching capture_temporal or capture_causal action even without an explicit request to save it.
  • Session persistence — tkg_id conversations now save session and message history on the server, the same way collection_id conversations already do. Pass the returned session_id back on your next request to continue the conversation with full context.
  • Consistent language — responses, follow-up questions, and suggested action titles now consistently match the language you’re writing in.
collection_id grounding is unchanged.
​
2026-07-16 · chat, TKG graph editing, fixes

​
Added: POST /api/v1/chat

A new streaming (SSE) chat endpoint.
  • Pass collection_id to ground answers in a collection’s knowledge (semantic chunk search with knowledge-graph fallback). Session and message history are persisted.
  • Pass tkg_id to ground answers in a standalone knowledge graph instead.
  • Omit both for a one-off chat with no grounding and no saved session.
  • Choose a model: gpt-5-nano, gpt-5-mini, or gpt-5.
  • Enable use_fact_search / use_evidence_finder / use_fact_checker to let the model autonomously call those research tools during the conversation. Each tool call is billed individually at the same credit cost as its standalone endpoint, only when actually invoked.
  • The chat request itself is currently free while pricing for this endpoint is being finalized.

​
Added: TKG graph editing endpoints

POST/PATCH/DELETE /api/v1/tkgs/{tkgId}/graph/nodes and POST/DELETE /api/v1/tkgs/{tkgId}/graph/edges let you manually create, rename, and delete nodes and edges on a knowledge graph that isn’t attached to a collection — the same graph-editing capability collections already had, now available standalone. These endpoints are free.A new PATCH /api/v1/tkgs/{tkgId}/graph/nodes/{nodeId}/annotation endpoint lets you attach a temporal annotation (when + precision: year/quarter/month/day) to a belief node.

​
Fixed: qa-verify entity matching

POST /api/v1/collections/{id}/qa-verify no longer over-matches on brand or product-family tokens that are common across many entities in the same collection (for example, matching almost any entity just because it shares a common brand name). Matching is now more precise.

​
Fixed: entity-connected fact limit

Raised the internal cap on how many connected facts are considered when resolving an entity from 20 to 50. Previously, some genuinely connected facts were cut off before relevance ranking, which could make a verdict appear unsupported when better evidence existed.
​
2026-07-14 · fact-search modes
GET /api/v1/fact-search now accepts a mode parameter:
  • auto (default) — semantic hybrid search with live fallback. This is the existing default behavior.
  • fast — keyword-only search that skips query embedding and the live fallback for the lowest latency. Best when you want the quickest results and don’t need semantic matching.
Omitting mode keeps the existing behavior (auto).

​
Improved: results when filtering by domain or narrow criteria

Fact-search now reliably returns matching articles when you restrict results (for example, by domain). Previously, narrow filters could return few or no results even when matching articles existed.
​
2026-07-08 · Faster qa-verify

​
Improved: faster response time

POST /api/v1/collections/{id}/qa-verify now resolves entity-linked evidence significantly faster per question/answer pair, reducing overall response time for the endpoint. No request or response shape changed.
​
2026-07-08 · Q&A verification improvements

​
Improved: Q&A verification evidence retrieval

POST /api/v1/collections/{id}/qa-verify now finds relevant evidence more reliably:
  • Evidence retrieval considers the claimed answer text as well as the question, so facts that only overlap with wording in the answer are no longer missed.
  • Entity matching now recognizes a question that mentions only part of an entity’s full name (e.g. a question naming a product’s short name is matched against the entity’s full name in the graph).
  • Vector search results are now filtered to a minimum relevance threshold, so clearly unrelated content is no longer included as evidence.

​
Fixed: one failing pair no longer fails the whole request

Previously, if any single question/answer pair in a qa-verify request hit an internal error, the entire request failed and no results were returned — even for pairs that would have verified successfully. Each pair is now processed independently: a failing pair now returns as "verdict": "unsupported" while the rest of the batch completes normally.No request or response shape changed.
​
2026-07-08 · Large collection graph fix

​
Fixed: 500 error on large collections

GET /api/v1/collections/{id}/graph could return a 500 error for collections with a few hundred or more claims/predictions/entities combined. Fetching a new collection’s graph for the first time after it grew past this size would fail entirely.Ingesting a new source into an already-large collection could also fail for the same underlying reason, blocking further ingestion into that collection.

​
What changed

Internal lookups that resolve node titles now run as a single server-side query instead of passing large ID lists in the request. This removes the size ceiling — collections with any number of nodes now work the same way.No response shape changed; if you were affected by this, requests that previously failed now simply succeed.
​
2026-07-07 · Entities in Collections

​
New: Entities in the knowledge graph

Collections now extract named entities (people, organizations, products) from ingested content, in addition to claims and predictions. Entities become nodes in the collection’s knowledge graph, connected to every claim or prediction that mentions them — turning previously disconnected facts about the same person, company, or product into a single connected cluster.

​
Entity types

​
Changed endpoints

​
Why this matters

Previously, retrieval relied only on vector similarity between the question and ingested text. If a question mentioned an entity by name but phrased things differently from the source content, relevant claims could be missed. Entities give the graph a second, exact-match retrieval path alongside semantic search.