majiayu000 Smart Sourcing
Intelligent sourcing that requires citations only when the cost is justified by the value of verification.
Philosophy
Full sourcing is prohibitively expensive (10-16x token increase). Smart sourcing targets high-value claims where verification materially improves accuracy.
When to Source
REQUIRE Sources
| Claim Type | Example | Why Source |
|---|---|---|
| Version numbers | "Python 3.12 added..." | Versions change, easy to verify |
| Performance claims | "30% faster than..." | Quantitative claims need evidence |
| Security recommendations | "Use bcrypt for..." | Security advice must be current |
| API specifications | "The function accepts..." | APIs change between versions |
| Release dates | "Released in Q4 2025" | Factual, verifiable |
| Pricing/limits | "Free tier allows 1000 requests" | Business terms change |
| Deprecated features | "X was removed in version Y" | Breaking changes need verification |
DO NOT Require Sources
| Claim Type | Example | Why No Source |
|---|---|---|
| General concepts | "Async improves concurrency" | Foundational knowledge |
| Code examples | Demonstrative snippets | Illustrative, not factual claims |
| Opinion/preference | "Consider using..." | Clearly framed as suggestion |
| Common knowledge | "Git tracks changes" | Universal understanding |
| Logical derivations | "Therefore, X implies Y" | Reasoning, not factual claim |
| Syntax | "Use def for functions" |
Language documentation is stable |
Decision Framework
Is this a factual claim?
├── No → No source needed
└── Yes → Would being wrong cause harm?
├── No (minor inconvenience) → No source needed
└── Yes → Is verification cost-effective?
├── No (>5k tokens for low-value claim) → State uncertainty
└── Yes → Perform web search, cite source
Implementation
Inline Citations
When sourcing is warranted:
Python 3.12 introduced `type` statement for type aliases[1].
Sources:
1. [PEP 695](https://peps.python.org/pep-0695/)
Uncertainty Markers
When sourcing isn't cost-effective but uncertainty exists:
Python 3.12 introduced the `type` statement (verify in release notes for your version).
Confidence Levels
| Confidence | Action |
|---|---|
| High (stable facts) | State directly |
| Medium (may have changed) | Add verification note |
| Low (uncertain) | Source or acknowledge uncertainty |
Token Cost Analysis
| Approach | Token Cost | Use Case |
|---|---|---|
| No sourcing | Baseline | General knowledge |
| Uncertainty marker | +5-10 tokens | Medium confidence |
| Single source | +2-5k tokens | High-value claim |
| Multiple sources | +10-20k tokens | Research tasks |
Examples
Good: Source Version-Specific Claim
User: What's new in FastAPI 0.109?
Response: FastAPI 0.109 added Pydantic v2 support as default[1] and
improved OpenAPI schema generation[2].
Sources:
1. [FastAPI Release Notes](https://fastapi.tiangolo.com/release-notes/)
2. [GitHub Release 0.109.0](https://github.com/tiangolo/fastapi/releases/tag/0.109.0)
Good: No Source for General Concept
User: How does async/await work?
Response: Async/await enables non-blocking I/O by allowing coroutines
to yield control while waiting for operations. The event loop manages
execution, resuming coroutines when their awaited operations complete.
Good: Uncertainty Marker
The connection pool default is typically 10 connections (verify in your
database driver documentation as this varies by library and version).
Integration
Works with:
conserve:response-compression- Keeps sourced responses conciseconserve:token-conservation- Weighs source cost vs valuememory-palace:research- Full sourcing for knowledge corpus
When To Use Full Sourcing
When NOT To Use
- Internal project code that doesn't need citations
- Casual conversation without factual claims
Escalate to full sourcing (accept high token cost) for:
- Knowledge corpus entries (permanent documentation)
- Security advisories (safety-critical)
- Compliance/legal claims (audit requirements)
- Research tasks (user expects thorough investigation)
For these cases, use memory-palace:research workflow which is designed for comprehensive sourcing.