The per-entity context layer your agents pull at runtime
Every agent that answers about a company, person or product needs the same thing first: what is actually true about this entity, right now, with sources. The Grounding Packet is that — one small, deterministic, signed payload of verified facts and corrections, fetched at the top of the workflow. Your first call, on every workflow.
ChatGPT, Claude, Gemini, Perplexity and Grokeach hold a slightly different — and slightly wrong — picture of most entities. Document RAG doesn’t fix identity facts: retrieval surfaces prose, not a verified value with a date and a source. The Grounding Packet closes that gap with a payload small enough to prepend to any prompt: verified facts, one per canonical field, each carrying source, observed_at-style dating and confidence — plus the corrections layer: the exact claims engines currently get wrong about the entity, with the verified value to prefer.
Grounding, not enrichment
Enrichment appends attributes. Grounding anchors answers. The difference is everything you need when an agent’s output has to stand up to scrutiny:
Provenance on every fact
Every value carries its source, observation date and confidence. Nothing in the packet is asserted without evidence — an agent can cite, not just claim.
Signed and verifiable
Ed25519-signed provenance (did:web:entidex.com). Recompute the content hash, resolve the DID document, verify the signature — a tampered packet fails.
Deterministic bytes
The same record always projects to an identical payload — stable ordering, one value per canonical field, no model in the loop. Safe to diff in CI and cache confidently.
Drift-monitored corrections
The packet tells you what AI engines get wrong about the entity today — the corrections layer is recomputed as engines drift, so the packet stays current when the record moves.
One call, three shapes
Fetch the packet with a free API key. The default shape is the canonical packet; append ?format=grounding-metadata for a groundingChunks / groundingSupports projection that drops into pipelines built against Google Vertex grounding shapes. Agents on MCP call the entity_grounding_packet tool instead.
# REST (free key)
curl -H "Authorization: Bearer $ENTIDEX_API_KEY" \
https://entidex.com/api/v1/entities/tesla/grounding-packet
# Vertex-shaped projection
curl -H "Authorization: Bearer $ENTIDEX_API_KEY" \
"https://entidex.com/api/v1/entities/tesla/grounding-packet?format=grounding-metadata"
# TypeScript SDK
const packet = await entidex.groundingPacket('tesla');The packet is the developer packaging of the Entity Knowledge Statement — the crawlable statement itself stays at /entity/{slug}/knowledge (markdown, JSON-LD, JSON), and the full statement API at /api/v1/entities/{id}/ground-truth. Same record, same signature — the packet is the shape built for agents.
Where it sits in an agent workflow
The pattern is fetch first, answer second. Before an agent asserts anything about an entity — a due-diligence summary, a sales brief, a support answer, a comparison — it pulls the packet and prefers the dated, cited values over whatever its base model remembers. The corrections layer does the heavy lifting: it is precisely the list of fields where model memory and the verified record currently diverge, so the agent knows exactly where its priors are stale.
The truth anchor
Packet values come from the verified record — registry data, Wikidata, platform metadata and the entity’s own published sources, collapsed to one value per canonical field. When sources genuinely disagree, that divergence is monitored and surfaced, never hand-edited away. The packet is a set of assertions to reconcile against, with the evidence attached — not a directive that binds any model.
Go deeper with Entidex
The full entity intelligence platform — beyond Explore Entidex
- Continuous multi-source entity observation
- Alerts when sources diverge or drift
- Cross-surface consensus + visibility over time
- Evidence-anchored intelligence reports
Frequently asked questions
What is a Grounding Packet?
A Grounding Packet is one small, deterministic per-entity payload of verified facts — each carrying its source, observation date and confidence — plus explicit corrections to claims AI engines commonly get wrong about the entity, and Ed25519-signed provenance. It is the developer packaging of the Entity Knowledge Statement: instead of scraping or guessing what is true about a company, person or product, an agent fetches the packet at the top of the workflow and answers against the verified record.
How is this different from entity enrichment?
Enrichment appends attributes; grounding anchors answers. An enrichment record is a snapshot with no provenance, no signature, and no monitoring — you cannot tell where a value came from, whether it is still true, or whether AI engines are contradicting it. A Grounding Packet carries per-fact provenance (source, observation date, confidence), a cryptographic signature you can verify against the issuer DID, and the drift-monitored corrections layer: the exact fields engines currently get wrong. Grounding, not enrichment.
What does "deterministic" mean here?
The same underlying record always projects to an identical payload — facts sorted stably, one value per canonical field, no sampling, no model in the loop. Determinism is what makes the packet safe to diff in CI, cache confidently, and reason about in tests: if the bytes changed, the record changed.
What is the grounding-metadata format?
Appending ?format=grounding-metadata returns the same record projected into a GroundingMetadata-shaped body — groundingChunks and groundingSupports over a synthesized fact digest, mirroring the field names used by Google Vertex grounding responses. Pipelines already built to consume that shape can ingest the packet without adapter code.
What does it cost?
The Grounding Packet is on the free API tier — sign up, create a free key, and call GET /api/v1/entities/{entityId}/grounding-packet. Calls are rate-limited per key. The packet is a pure projection over data the platform already verifies and stores, so there is no per-call model cost inflating the price.
How do I verify the signature?
When signing is configured the packet carries provenance.proof — an Ed25519 JWS over the content hash of the underlying statement, issued as did:web:entidex.com. Resolve the DID document at entidex.com/.well-known/did.json, recompute the content hash, and verify the signature. A tampered or stale packet fails verification.