Skip to content

Config reference: hydration

  • Enterprise tuning surface

    Defaults + constraints are rendered directly from Pydantic.

  • Env keys when available

    Many fields have an env-style alias (from TriBridConfig.to_flat_dict()).

  • Tooltip-level guidance

    If a matching glossary entry exists, you’ll see deeper tuning notes.

Config reference Config API & workflow Glossary

Total parameters: 2

Group index
  • (root)

(root)

JSON key Env key(s) Type Default Constraints Summary
hydration.hydration_max_chars HYDRATION_MAX_CHARS int 2000 ≥ 500, ≤ 10000 Max characters to hydrate
hydration.hydration_mode HYDRATION_MODE str "lazy" pattern=^(lazy|eager|none|off)$ Context hydration mode

Details (glossary)

hydration.hydration_max_chars (HYDRATION_MAX_CHARS) — Hydration Max Chars

Category: general

HYDRATION_MAX_CHARS limits how much raw chunk text is loaded when turning retrieval hits into generation-ready context. This protects latency, memory, and prompt budgets from oversized chunks or unusually large files. A value that is too low can cut away critical lines and reduce answer grounding, while a value that is too high can bloat prompts and increase cost without better relevance. Tune this alongside chunk size, rerank depth, and model context window so each stage has a clear budget. Track truncation rate and citation failures to detect when the cap is suppressing necessary evidence.

Badges: - Prompt Budget

Links: - LangChain Contextual Compression - LangChain Text Splitters Concepts - OpenAI Cookbook: Count Tokens with tiktoken - MoC (2025): Mixtures of Chunking Learners for RAG

hydration.hydration_mode (HYDRATION_MODE) — Hydration Mode

Category: general

HYDRATION_MODE controls when full chunk content is loaded into the retrieval pipeline. A lazy mode usually gives the best production balance because ranking can run on lightweight metadata first, and full text is fetched only for shortlisted results. A none mode is useful for retrieval diagnostics, fast metadata-only workflows, or systems where generation occurs in a separate step. Choosing the mode changes both latency profile and memory behavior, so it should be treated as an architectural runtime option rather than a cosmetic toggle. Align this setting with your reranker strategy and context assembly stage to avoid unnecessary data loading.

Badges: - Runtime Loading

Links: - LangChain Contextual Compression - LlamaIndex Node Parsers - OpenAI Cookbook: Count Tokens with tiktoken - TeleRAG (2025): Lookahead Retrieval for Efficient Inference