> ## Documentation Index
> Fetch the complete documentation index at: https://docs.graphorlm.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Model Use and Training

> How Graphor uses AI models, what customer data reaches which provider, and the contractual commitments that prevent customer content from being used to train any model.

## The commitment

**Customer Content sent to Graphor is never used to train any AI model — neither Graphor's own models nor the models of any subprocessor in the inference chain.**

This is a contractual commitment, not just a best-effort promise. It rests on three layers:

1. **Synapse Terms of Service** — "Synapse does not train AI models using User Content."
2. **Per-provider contractual no-training clauses** — every AI subprocessor in the Graphor production chain (AWS Bedrock, OpenAI, Cerebras) has a published no-training policy that Graphor relies on by contract. Verbatim citations are in [§4 below](#4-per-provider-evidence-verbatim).
3. **Operational controls** — Graphor does not opt in to any provider's training programs, does not maintain a fine-tuning pipeline on customer corpora, and routes traffic only through provider tiers with the no-training default.

This page is the canonical reference for that commitment. The [Subprocessors](/trust/subprocessors) page lists who sees customer data; this page documents what they are allowed to do with it.

## 1. What gets to an AI provider

Two distinct moments send Customer Content (or content derived from it) to an external AI provider. Both are governed by the no-training commitment above.

### 1.1 Indexing time (once per build)

When you ingest a new source (or rebuild an existing one), Graphor produces an indexed retrieval representation. This involves two AI providers:

| Step             | Provider                              | What is sent                                                                                                | What comes back                    |
| ---------------- | ------------------------------------- | ----------------------------------------------------------------------------------------------------------- | ---------------------------------- |
| Chunk enrichment | **Cerebras** (`gpt-oss-120b`)         | Chunked text from your source + per-page and per-document metadata annotations to be appended to the chunk. | Enriched chunk text for embedding. |
| Embedding        | **OpenAI** (`text-embedding-3-small`) | The enriched chunk text.                                                                                    | A 1 536-dimensional vector.        |

The vectors are stored in Graphor's own managed graph store. After embedding, **no further indexing-time traffic to either provider occurs for that build**.

### 1.2 Query time (per `sources.ask` or `/data-extraction` request)

When a user asks a question or runs a structured extraction, Graphor routes the request to one of two **tiers** based on the `thinking_level` parameter the caller provides.

| `thinking_level`                        | Tier              | Provider chain                                                                                                                                                                            | Latency profile                                                                                     |
| --------------------------------------- | ----------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |
| `fast`                                  | **Fast tier**     | **Cerebras** (`gpt-oss-120b`) inference. If the Cerebras path is unavailable, the request falls back to the Standard tier provider.                                                       | Sub-second answers for short queries.                                                               |
| `balanced`, `accurate` (default), `max` | **Standard tier** | **Anthropic Claude** family, hosted on **AWS Bedrock**. The `thinking_level` controls the reasoning effort applied to the same model, from medium effort up to maximum extended thinking. | Several seconds for `balanced`; tens of seconds for `accurate` and `max` on long-context reasoning. |

What is sent on each request:

* The user's question text
* A bounded set of context retrieved from your indexed sources (relevant snippets, with provenance metadata)
* The system prompt and conversation history for that session
* For `/data-extraction` requests: the JSON Schema describing the expected output shape

What comes back: the model's textual answer (or structured extraction matching the schema), plus token usage metadata for billing and observability.

**Customer Content does not leave the Graphor production environment** except as required to fulfill one of these two flows. There is no batch upload to provider training endpoints, no nightly export, and no cross-customer aggregation.

## 2. Customer-controllable knobs

The following parameters and project-level settings give the customer direct control over what touches an AI provider and how:

| Knob                  | Where                                               | Default                                                                                                             | What it controls                                                                                                                                                                                        |
| --------------------- | --------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `thinking_level`      | Per request on `sources.ask` and `/data-extraction` | `accurate` (Standard tier)                                                                                          | Routes the request to the Fast or Standard tier as described in [§1.2](#12-query-time-per-sourcesask-or-data-extraction-request). Set `fast` for low-latency answers, `max` for long-context reasoning. |
| Reranker              | Project setting                                     | **Off**                                                                                                             | When enabled, applies an OpenAI reranking pass to retrieved chunks before final answer composition. Off by default — when off, OpenAI is not invoked at query time.                                     |
| Hybrid search         | Project setting                                     | **Off**                                                                                                             | When enabled, combines vector and full-text search (no LLM, no additional AI subprocessor involvement).                                                                                                 |
| Observability tracing | Project setting (tier-aware default)                | **Off for Enterprise tier projects; On for Free/Pro tier projects** with the Brazilian PII mask applied before send | When enabled, prompts and completions are sent to Graphor's self-hosted observability store. See [Tenant Isolation](/trust/tenant-isolation) and [Data Retention](/trust/data-retention).               |

The customer cannot opt **in** to any model-training program through Graphor's API — there is no flag, parameter, or setting that would cause Customer Content to be used for training.

## 3. What Graphor uses Customer Content for

For transparency, here is the complete enumeration of how Customer Content is used inside the Graphor production environment:

1. **Storage** — the raw source you upload is stored in Graphor's project-scoped object storage in `us-central1` for as long as the source exists in your project. Deleted on customer DSR action per [Data Retention](/trust/data-retention).
2. **Partitioning** — the source is parsed into structured elements (pages, sections, tables, images) by Graphor's document parsing service, running on Graphor's own infrastructure.
3. **Chunking** — partitioned text is split into retrieval-sized chunks by Graphor's own code.
4. **Enrichment** — chunks are annotated with per-page and per-document context using the Cerebras-hosted enrichment model. The enriched text is what gets embedded.
5. **Embedding** — chunks are sent to OpenAI for embedding under the OpenAI Zero Data Retention agreement (no logging, no training).
6. **Indexing** — vectors and metadata are written to Graphor's own managed graph store.
7. **Retrieval** — at query time, Graphor's own code searches the index and selects relevant chunks.
8. **Inference** — selected chunks plus the user's question are sent to either the Fast or Standard tier provider for an answer.
9. **Operational improvement** — anonymized, mask-applied operational telemetry (latencies, error rates, retrieval quality signals) may be used by Synapse to improve the Graphor product. **This explicitly excludes the content of customer questions and answers.**

Items 4, 5, and 8 are the only steps in which an external AI provider receives Customer Content. All three providers are governed by the no-training commitments cited verbatim below.

## 4. Per-provider evidence (verbatim)

Each provider's public position on training and retention is reproduced here exactly as published. Sources are linked so any reader can verify.

### 4.1 Anthropic, PBC — model owner of Claude

Anthropic models reach the Graphor pipeline through AWS Bedrock; **no Customer Content transits Anthropic's own infrastructure**. The model-owner commitment from Anthropic's own commercial terms reinforces the AWS-side commitment in [§4.2](#42-amazon-web-services-aws-bedrock):

> "Anthropic may not train models on Customer Content from Services."

Source: [Anthropic Commercial Terms of Service](https://www.anthropic.com/legal/commercial-terms), Section B (Customer Content).

### 4.2 Amazon Web Services — AWS Bedrock

Bedrock is the production lane for every Standard-tier and fast-tier-fallback inference call:

> "No, AWS and the third-party model providers will not use any inputs to or outputs from Amazon Bedrock to train Amazon Nova, Amazon Titan, or any third-party models."

> "No. Users' inputs and model outputs are not shared with any model providers."

> "Any customer content processed by Amazon Bedrock is encrypted and stored at rest in the AWS Region where you are using Amazon Bedrock."

Source: [AWS Bedrock FAQs](https://aws.amazon.com/bedrock/faqs/), Security section.

### 4.3 OpenAI, L.L.C. — embeddings and optional reranker

OpenAI provides the embedding model used during indexing and (when explicitly enabled) the reranker:

> "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)."

Source: [OpenAI API — Your Data](https://developers.openai.com/api/docs/guides/your-data), Data controls section.

Additionally, **Graphor is enrolled in OpenAI Zero Data Retention (ZDR)**, which eliminates the default 30-day abuse-monitoring log for API requests. ZDR enrollment status is audited by Synapse at least quarterly. If ZDR is ever disabled (whether by Synapse choice or by OpenAI policy change), customers under a [Data Processing Addendum](/legal/dpa-template) are notified at least 30 days before any planned disablement, or within 24 hours of becoming aware of an unplanned OpenAI-side disablement, per [DPA §4.4](/legal/dpa-template#44-subprocessor-obligations-and-pass-through-clauses); this page is updated and subscribers to [subprocessors@graphorlm.com](mailto:subprocessors@graphorlm.com) are also notified.

### 4.4 Cerebras Systems, Inc. — chunk enrichment and fast tier

Cerebras serves the open-source `gpt-oss-120b` model used for chunk enrichment and the fast tier:

> "the foregoing does not grant Cerebras the right to use Service Content for the purpose of training or fine-tuning models."

Source: [Cerebras Terms of Use](https://www.cerebras.ai/terms-of-service).

Cerebras additionally commits to zero retention of inference data:

> "We do not retain inputs and outputs associated with our training, inference and chatbot Services."

Source: [Cerebras Privacy Policy](https://www.cerebras.ai/privacy-policy).

## 5. Currently running

The tier-based declaration in [§1.2](#12-query-time-per-sourcesask-or-data-extraction-request) is the stable contract — provider and tier do not change without a published update to this page. The exact model SKUs that serve each tier do change as providers release new versions; the table below records the models in production as of the **Last updated** date in the front matter.

| Step                                                    | Currently running                                                                             | Effective from |
| ------------------------------------------------------- | --------------------------------------------------------------------------------------------- | -------------- |
| Chunk enrichment                                        | `gpt-oss-120b` on Cerebras                                                                    | 2026-06-21     |
| Embedding                                               | `text-embedding-3-small` on OpenAI                                                            | 2026-06-21     |
| Fast tier inference                                     | `gpt-oss-120b` on Cerebras (fallback: Claude Haiku on AWS Bedrock if Cerebras is unavailable) | 2026-06-21     |
| Standard tier inference (`balanced`, `accurate`, `max`) | Anthropic Claude Opus on AWS Bedrock                                                          | 2026-06-21     |
| Reranker (when enabled by the customer)                 | OpenAI reranking model                                                                        | 2026-06-21     |

Model upgrades within the same tier (for example, a future Claude Opus version) are deployed as part of normal release management; the table above is revised on each upgrade and the [change history](#7-change-history) records the date.

## 6. What Graphor explicitly does NOT do

Stating these explicitly removes ambiguity for security and privacy reviewers:

* **No fine-tuning on customer corpora.** Graphor does not maintain a pipeline that fine-tunes any model on customer-derived datasets.
* **No opt-in to provider training programs.** Synapse does not opt in to any provider's training programs (for example, OpenAI's training opt-in) for the production Graphor org.
* **No cross-customer data aggregation for model improvement.** Operational telemetry used to improve the Graphor product is aggregated and mask-applied; the content of customer questions and answers is not part of it.
* **No human review of customer prompts for quality grading.** Synapse personnel do not review customer Content for quality grading, dataset curation, or evaluation purposes.
* **No silent provider substitution.** Provider and tier changes are published in advance via the [Subprocessors](/trust/subprocessors) page and the `subprocessors@graphorlm.com` notification list.

## 7. Change history

| Version | Date       | Change               |
| ------- | ---------- | -------------------- |
| 1.0     | 2026-06-21 | Initial publication. |

When the tier-to-provider mapping, the per-provider citation, or the currently-running model annex changes, this table is updated and subscribers to [subprocessors@graphorlm.com](mailto:subprocessors@graphorlm.com) receive an email.

## Contact

* General privacy and DPA inquiries: [privacy@graphorlm.com](mailto:privacy@graphorlm.com)
* Subprocessor and model-change notifications: [subprocessors@graphorlm.com](mailto:subprocessors@graphorlm.com)
* Customer support: [support@graphorlm.com](mailto:support@graphorlm.com)