Context Engine

Every answered query becomes a searchable knowledge asset. The Context Engine extracts atomic facts from LLM responses, embeds them, stores them in a vector database, and retrieves them to enrich future queries. The more queries flow through Xilos, the richer the knowledge base becomes.

How It Works

The Context Engine operates in three phases:

1. Extract — After a query completes, Xilos parses the response and decomposes it into atomic facts: self-contained, verifiable statements.
2. Embed — Each fact is embedded with the local embedding model and stored in the vector database alongside metadata (source query, model, timestamp, quality score).
3. Retrieve — When a new query arrives, Xilos embeds the query, searches the fact store for semantically similar facts, and injects them into the prompt as context before the LLM is called.

Info: The Context Engine is what makes Xilos compound knowledge over time. A query answered today enriches the response to a related query asked weeks later — even if phrased entirely differently.

Atomic Facts

An atomic fact is a single, self-contained statement extracted from a response. Decomposing responses into atomic facts (rather than storing whole responses) improves retrieval precision and reduces noise.

Example

Given the response:

"Xilos was founded in 2024 and is headquartered in San Francisco. The platform supports Claude, GPT, and Gemini model families, with semantic caching that reduces costs by up to 90%."

The Context Engine extracts four atomic facts:

1. Xilos was founded in 2024.
2. Xilos is headquartered in San Francisco.
3. The Xilos platform supports Claude, GPT, and Gemini model families.
4. Xilos semantic caching reduces costs by up to 90%.

Each fact is embedded and stored independently. A future query about "where is Xilos based" retrieves fact #2, while a query about "supported AI models" retrieves fact #3 — without pulling in unrelated information.

Fact Extraction

Extraction runs automatically after every completed query. The pipeline:

Parse response

The response text is split into sentences and clauses.

Decompose

An extraction model breaks compound statements into atomic facts.

Deduplicate

New facts are compared against existing facts. Duplicates and near-duplicates are skipped.

Score

Each fact receives a faithfulness score based on the source response's quality metrics. Low-scoring facts are flagged.

Store

The fact is embedded and written to the vector database with full provenance metadata.

Warning: Facts are only extracted from responses that pass quality thresholds. If a response has a low faithfulness score, its facts are quarantined and excluded from retrieval until reviewed.

Semantic Retrieval

When a new query arrives, the Context Engine retrieves relevant facts before the LLM is called:

1. The query is embedded with a local model.
2. Xilos performs a cosine similarity search against the fact store.
3. The top-N facts above the similarity threshold are selected.
4. Selected facts are injected into the prompt as a context block.

Retrieved facts appear in the Query Log alongside the response, so you can inspect exactly which facts informed each answer.

Retrieval Configuration

Document Ingestion

Beyond query responses, you can ingest external documents directly into the Context Engine. Uploaded documents are processed through the same extraction pipeline: parsed, decomposed into atomic facts, embedded, and stored.

Supported Formats

• Plain text (.txt, .md)
• PDF (.pdf)
• Word documents (.docx)
• HTML (.html)

Ingesting a Document

Dashboard

1. Navigate to Context Engine → Documents.
2. Click Upload Document.
3. Select a file or paste text.
4. Optionally tag the document with metadata (source, category, expiry).
5. Click Ingest. Xilos processes the document and reports the number of facts extracted.

API

curl -X POST https://api.xilos.ai/v1/context/documents \
  -H "Authorization: Bearer $XILOS_API_KEY" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@report.pdf" \
  -F "category=financial" \
  -F "source=Q3-2024-Report"

Info: Large documents are chunked before extraction. Each chunk is processed independently, and the resulting facts are deduplicated across chunks before storage.

Configuration

Configure the Context Engine at the organization level in Settings → Context Engine.

Extraction Settings

Embedding Settings

Storage Settings

Privacy and Data Isolation

• Facts are scoped to the organization. They are never shared across tenants.
• Per-rule overrides can disable fact extraction for sensitive queries (e.g., PII handling).
• Ingested documents can be tagged with an expiry, after which their facts are automatically purged.

Warning: If a routing rule handles sensitive or regulated data, disable Context Engine extraction for that rule to prevent storing facts derived from confidential information.

Best Practices

• Ingest foundational documents — Upload policies, product specs, and FAQs so the Context Engine has a strong knowledge base from day one.
• Tune the similarity threshold — Lower it for broader recall; raise it for stricter precision.
• Monitor fact quality — Review low-faithfulness facts in the Context Engine dashboard before they pollute retrieval.
• Set expiries on time-sensitive documents — Use fact_ttl_days or document-level expiry tags to auto-purge stale facts.
• Disable extraction for sensitive rules — Protect PII and confidential data by disabling extraction per-rule.