Prompts

The Prompts table stores text templates that Vedana uses across the pipeline. The name column is the key that the code looks up the template by. If a row for the name isn’t found, the built-in default is used (see vedana_core.llm and vedana_core.data_model).

Fields

FieldDescription
nameTemplate name (key).
textTemplate body (with placeholders).

Available templates

KeyWhere it’s usedPlaceholders
generate_answer_with_tools_tmpltMain system prompt for the agent (LLM.generate_cypher_query_with_tools){tools}, {graph_description}
dm_filter_promptSystem prompt for the data model filtering step—
dm_filter_user_promptUser prompt for the data model filtering step{user_query}, {compact_data_model}
finalize_answer_tmpltSystem prompt for finalisation after the tool-calling loop is exhausted—
generate_no_answer_tmpltPrompt for the “nothing found” scenario—
dm_descr_templateTemplate that renders the entire data model{anchors}, {anchor_attrs}, {links}, {link_attrs}, {queries}
dm_anchor_descr_templateTemplate for one anchor row{anchor.noun}, {anchor.description}, {anchor.id_example}, {anchor.query}
dm_attr_descr_templateTemplate for one attribute row{anchor.noun}, {attr.name}, {attr.description}, {attr.example}, {attr.query}
dm_link_descr_templateTemplate for one link row{link.sentence}, {link.description}, {link.query}
dm_link_attr_descr_templateTemplate for one link attribute row{link.sentence}, {attr.name}, {attr.description}, {attr.example}, {attr.query}
dm_query_descr_templateTemplate for one Queries row{query.name}, {query.example}

Defaults

Every template has a default hardcoded in the code. That means you can run Vedana without entries in Prompts — it’ll just work with the built-in prompts (in Russian by default for Vedana’s defaults).

Defaults:

  • generate_answer_with_tools_tmplt — the main system prompt for the agent; see vedana_core/llm.py.
  • dm_filter_* — prompts for data model filtering; see vedana_core/rag_pipeline.py.
  • finalize_answer_tmplt — finalisation.
  • generate_no_answer_tmplt — “nothing found” message.
  • dm_*_descr_template — render the data model into human-readable sections.

The full text of the defaults is in the source:

  • libs/vedana-core/src/vedana_core/llm.py
  • libs/vedana-core/src/vedana_core/rag_pipeline.py
  • libs/vedana-core/src/vedana_core/data_model.py

When to override

Main system prompt (generate_answer_with_tools_tmplt)

The most common — override for:

  • tone and style of the assistant (formal/friendly, formal “you” / informal “you”);
  • domain vocabulary: “we issue orders”, not “we hand over goods”;
  • hard rules: “never answer about competitor prices”, “decline when personal data is mentioned”;
  • answer format: always return a list, always include a link, always mention the ID.

Filtering prompt (dm_filter_prompt / dm_filter_user_prompt)

Rarely overridden, but sometimes useful:

  • if you have a special structure of model description;
  • if you want the model to be more aggressive or more lenient in selecting attributes.

Finalisation (finalize_answer_tmplt)

Triggers when the tool-call iteration limit is reached. Worth overriding to:

  • give an honest “I couldn’t get an answer, here’s what I found” with the limitations.

No answer (generate_no_answer_tmplt)

When nothing was found. Worth overriding to:

  • propose relevant alternative questions;
  • give a contact point (support, feedback form);
  • avoid over-apologising — sounds fake.

Data model rendering templates

Override if you want:

  • to shrink the context (drop query for attributes);
  • to add more structure (markdown tables instead of lists);
  • to localise section headers.

How to update

  1. Open Grist > Data Model > Prompts.
  2. Create or update a row with the right name.
  3. In the backoffice → ETL run data_model_steps (or “Refresh Data Model”).
  4. Run an evaluation to confirm metrics didn’t drop.

Changes take effect immediately after refresh.

Best practices

  • Don’t write a prompt-novel. The longer the system prompt, the more expensive every request is (it goes into every LLM call in the thread).
  • Clear “do / don’t” rules work better than vague guidelines.
  • List tools explicitly if you want to constrain choices: “Use only cypher; vector_text_search — only if cypher returns an empty result”.
  • Test prompts through evaluation, not “by feel” in chat — that gives you a hit-rate metric, not a subjective feeling.
  • Version: give the prompt a date/version suffix (vedana_main_v3) and keep older versions next to it so you can roll back.

Debugging a prompt

If something went wrong, in the backoffice → Chat → Details next to the answer you can see:

  • the final system prompt sent to the LLM (with {graph_description} and {tools} substituted);
  • the message set (including the dm_filter_* instructions if filtering is on);
  • the list of executed tool calls;
  • the finalisation step (if it was applied).

That helps you understand whether the LLM is really receiving what you wrote.

What NOT to change unnecessarily

  • Template key names. They are hard-coded in source; a wrong name means the code falls back to the default, not your text.
  • Placeholders ({tools}, {graph_description}, {user_query}, {compact_data_model}). Without them the templates won’t work.
  • Final “verify” instructions. If you want to check answer correctness, do it via evaluation — don’t bake “double-check yourself” into the prompt.

What’s next