Skip to main content
LCA Talents Hub

How Ask LCA works

Ask LCA is an AI assistant built on RAG. It answers questions about life cycle assessment (LCA), carbon footprints and information in the network. For every question, it first retrieves content from the LCA Hub website, then has an LLM answer from that content, with a citation for every fact.

  • Who can use itApproved members. Guests who are not signed in can browse the FAQ.
  • What it answers fromContent published on this website, and opt-in expert data
  • You can check itA citation such as [1] links to the original page

Steps

From question to answer

Every question goes through these 8 steps in order, in a few seconds. The dashed boxes show where data goes out to an AI provider.

  1. 1You type a question

    Ask in Thai or English, up to 1,000 characters.

  2. 2Sanitizer

    Tries to remove contact details from the question, such as phone numbers, email addresses and links. This is best effort and cannot catch every form, so please do not type personal data in your question.

  3. 3Contact guard

    Checks whether the question asks for a person's contact details, for example a speaker's phone number or email. No ⇒ go to the next step

    Yes ⇒ stop hereReplies with the decline sentence at once, does no retrieval and does not send the question to an AI provider.
  4. 4Retrieval with hybrid search

    Searches the index in two ways at once: lexical search finds chunks that contain the same words (a glossary matches Thai and English terms), and vector search finds chunks whose embedding is close to the question, even when the words differ. The two rankings are then merged by fusion (RRF).

    OpenAICreates the embedding of the question, for vector search.
  5. 5Access check on every source at answer time

    Uses only sources that are still published and that you are allowed to see, such as knowledge library items open to members only. Expert data is used only from experts whose opt-in is still active.

  6. 6Context selection

    Picks about 8 of the best-matching chunks from the retrieval results as the context. The LLM reads only this context, not the whole website.

  7. 7The LLM generates the answer

    The LLM follows the system prompt: answer only from the context, and put a citation on every fact. If the information is not enough, it must reply with the decline sentence “There is not enough information to answer this question.” instead of guessing (hallucination).

    Anthropic (Claude)The LLM runs in the United States. It receives the question, the latest chat history from the same chat, and the context.
  8. 8The answer with citations

    A list of sources follows the answer. Select a citation or a source name to open the original page, then press 👍 or 👎 to give feedback on whether the answer was helpful.

Source

What the LLM reads

Ask LCA does not search the internet, and does not mainly answer from what the LLM learned during training. The LLM reads only chunks from our own index. They come from 2 groups:

  • Content published on the website: the FAQ, the knowledge library, news, activities, projects and the project timeline
  • Opt-in expert data: profiles (name, organization, position, short bio, expertise), publications and resources. Experts choose for themselves whether to give their opt-in, and can withdraw it at any time. The system does not send contact details, personal websites or profile pictures to an AI provider.
Ingest: how content gets into the indexPublished content and opt-in expert data go through chunking, then OpenAI creates an embedding of each chunk, and the result is stored in the Ask LCA index. Ingest runs about every 5 minutes.SourcePublishedcontent andopt-in expertsChunkingthen createembedding(OpenAI)Indexof Ask LCAingest runs aboutevery 5 minutes
Ingest adds new or edited content to the index within about 5 minutes.
Unpublished ⇒ no longer used, at once

When an administrator unpublishes content, or an expert withdraws their opt-in, the next answer no longer uses that data, because there is an access check every time Ask LCA answers (step 5).

Members-only content

Knowledge library items open only to members are used to answer only members who are allowed to see that content on the website.

Answers

How to read an answer

An example of the layout only, not a real answer from the system

What is an organization's carbon footprint?

An organization's carbon footprint is the amount of greenhouse gases the organization emits from all of its activities, in tonnes of carbon dioxide equivalent [1]. It counts both direct emissions and emissions from the energy it buys and uses [2].

Sources
  1. An article in the knowledge library
  2. A frequently asked question
Citation

The numbers [1] [2] show which source a sentence comes from. Select one to open the original page. If that source has been withdrawn, the list shows that it is not available.

Decline sentence

“There is not enough information to answer this question.” means the index does not yet have information that answers this question. The system replies this way on purpose instead of guessing. Try a more specific question, or look at the list of experts.

The LLM can be wrong

Answers may be incorrect or incomplete and are not professional advice. Check them against the source or with an expert before you use them.

Chat history

The LLM sees the latest earlier messages in the same chat, so you can ask follow-up questions. If you change the topic, starting a new chat works better.

Background

Why RAG instead of fine-tuning

One way is training or fine-tuning an LLM so that it remembers the content in its model weights. The other way is RAG: the LLM reads content that is retrieved before every answer. Ask LCA uses RAG. The model itself is not changed; what we tune is the retrieval, the context selection and the system prompt.

Training / fine-tuningAsk LCA (RAG)
New contentNeeds another round of fine-tuning, which takes days to monthsIngest adds it to the index within about 5 minutes
Removing or hiding dataHard, because the data is already built into the model weightsStops being used at once when it is unpublished or the opt-in is withdrawn
Naming the sourceCannot tell where it learned somethingA citation on every fact
CostHundreds to millions of dollars per roundA few satang (hundredths of a baht) per question
When there is no informationRisk of hallucinationThe system prompt requires the decline sentence

Privacy

Where your data goes

A summary of the Ask LCA privacy notice. If anything on this page differs from the notice, the notice prevails.

How your data travels when you ask Ask LCAYou send a question to the LCA Hub website, which does not attach your name, email or phone number, and passes it on to three parties: Anthropic, in the United States, receives the question, the chat history in the same chat and context from the website, so that the LLM can generate the answer; OpenAI, which processes in several countries, receives the question and content from the website, to create embeddings for vector search; Supabase, in Singapore, is the database that stores chats, messages, usage data and consent records, and messages are deleted automatically after 180 days.You (member)type a questionLCA Hub websiteDoes not attach your name, email or phone numberAnthropic, PBCUnited StatesReceives: the question, the chat historyin the same chat, and context from the siteFor: the LLM to generate the answerOpenAI OpCo, LLCSeveral countriesReceives: the question and website contentFor: creating embeddingsfor vector searchSupabase, Inc.SingaporeStores: chats, messages, usage dataand your consent recordsMessages are deleted automaticallyafter 180 days
Anthropic and OpenAI are the AI providers; Supabase is the database of the website.
  • Consent The system asks for consent before your first use, and you can withdraw it at any time on the settings page. When you withdraw, the system stops sending your data to the AI providers and deletes your whole chat history from the main system immediately. The backup (about 14 days) and data already sent to the AI providers (no more than 30 days) are deleted on their own schedules.
  • Who sees your chats and your usage In the system, only you can see your chats. Site administrators have no screen for reading chat content. They see aggregate statistics and a monthly usage summary for each member, such as the number of questions and the estimated cost, with no chat content. The exceptions are fixing a system fault, investigating a security breach or complying with the law. The WAF may also record a request that was rejected or caused an error, for up to about 1 day. See clause 8 of the notice.
  • AI providers Under their current API terms, Anthropic and OpenAI do not use data sent through the API for training models by default, and keep it for no more than 30 days before deleting it, except where the provider requires otherwise or the law requires it.
  • We do not use your chats for training models Quality improvements use aggregate statistics and the 👍 👎 feedback you give.
  • Please do not type sensitive data such as health, religion, national ID numbers, passwords or other people's personal data. If you type it in a question yourself, it may be sent with the question.

All the details, including your rights under the personal data protection law, are in the Ask LCA privacy notice.

Usage

Daily credit

Each member has a daily credit. The credit bar at the top of the chat page shows how much is left today. Credit resets at midnight, Thailand time.

  • Members can usually ask about 15–30 questions a day. Long questions or answers use more credit (still at least about 7 questions).
  • Experts and administrators have more credit.
  • Everyone can ask no more than 30 questions a day.
  • Fallback when today's credit runs out: the system still shows the FAQ and links to related pages, but the LLM does not generate answers until the next day.

Quality

Evaluation before launch

Before opening Ask LCA to members, we built a golden set of questions whose correct answers we know in advance, currently 78. It has questions that our website can answer and questions it cannot (to see whether the LLM replies with the decline sentence). We then ran an evaluation against the real system and checked every answer. We tuned the system and ran the evaluation again over several rounds, until it passed every gate.

  • ≥ 9 in 10answers have correct facts
  • ≥ 95%of citations point to a source that really contains that text
  • Allquestions without information get the decline sentence
  • 0 timesprivacy leak (disclosing personal data that should not be disclosed)

After launch, we keep watching aggregate statistics and 👍 👎 feedback to improve, without reading members' chat content. Passing the gates does not mean it cannot be wrong, so always check against the source.

Tips

How to ask for a good answer

  • Ask about things likely to be on this website, such as LCA knowledge, activities, news, projects or experts in the network.
  • Be specific, for example with an event name, a year or a topic, rather than asking broadly, because retrieval then finds the right chunks more easily.
  • Ask about one topic at a time. If you change the topic, start a new chat.
  • Send your question with the send button or press Ctrl + Enter (on a Mac, ⌘ + Enter). Enter alone starts a new line.
  • To contact an expert, go to the experts page. The contact guard of Ask LCA does not give out people's contact details.

Technical terms

Technical terms on this page

Show all 44 terms (in the order they appear on this page)
AI
Artificial intelligence. On this page it means a program that can read and write language.
LLM (large language model)
An AI model trained on a huge amount of text until it can read and write language. Ask LCA uses Claude from Anthropic.
RAG (retrieval-augmented generation)
A way to have an LLM search for and read related content before it answers, instead of answering from memory.
Retrieve / retrieval
The step that searches the index for chunks related to the question.
Citation
The numbers [1] [2] in an answer. They show which source a fact comes from and link to the original page.
FAQ
Frequently asked questions, with answers that the site administrators have written.
Opt-in
The data owner chooses to allow it; the default is not allowed. Experts give their opt-in on their own settings page and can withdraw it at any time.
AI provider
A company that runs AI models for us: Anthropic (LLM) and OpenAI (embedding).
Sanitizer
A filter that removes patterns of contact details, such as phone numbers, email addresses and links, from text before it is passed on.
Best effort
Done as well as possible, with no guarantee that it catches 100%.
Contact guard
A check that catches requests for a person's contact details and replies with the decline sentence at once, without sending the question out.
Decline sentence
The fixed sentence “There is not enough information to answer this question.”, used when there is no answer, with no other text mixed in.
Index
The store of all chunks, with the embedding of each chunk, so that they can be searched quickly.
Chunk / chunking
Chunking cuts long content into short pieces; each piece is a chunk. A chunk is the unit that the system searches and sends to the LLM to read.
Embedding
A set of numbers that represents the meaning of a text. Texts with similar meanings get similar numbers, even when they use different words or languages.
Hybrid search
Searching with lexical search and vector search together, then merging the results.
Lexical search
Searching for chunks that contain the same words as the question.
Glossary (in search)
A list of Thai and English terms that mean the same thing, such as คาร์บอนฟุตพริ้นท์ and carbon footprint, so that lexical search finds both languages.
Vector search
Searching for the chunks whose embedding is closest to the embedding of the question, that is, searching by meaning.
Fusion (RRF)
Merging the rankings of the two searches into one list. Chunks that rank well in both come first. RRF stands for reciprocal rank fusion.
Access check
A check at answer time that a source is still published, that the person asking is allowed to see it, and that the expert's opt-in is still active.
Source
The original page that a chunk comes from, such as an article in the knowledge library, a news item or an activity.
Context / context selection
The context is the set of about 8 chunks sent to the LLM to read for one answer. Context selection is the step that picks this set.
Generate
The LLM writing out the text of the answer.
System prompt
Fixed rules that the LLM must follow in every answer, such as answering only from the context and adding citations.
Hallucination
When an LLM makes up information that does not exist and makes it sound believable.
Chat history
The latest earlier messages in the same chat, sent with the question so that you can ask follow-up questions.
Feedback
Pressing 👍 or 👎 under an answer. There is no field for typing text.
Training
Building an AI model from a huge amount of text. It takes months and millions of dollars.
Fine-tuning
Training an existing model further with specialist data, so that it remembers that content itself.
Model weights
The numbers inside a model that hold what it has learned. Training and fine-tuning change these numbers.
Ingest
An automatic job that adds new or edited content to the index. It runs about every 5 minutes.
Database
The system that stores the data of the LCA Hub website. It uses Supabase.
Usage data
Data about use, such as the date and time, the credit used, the amount of processing and the error status, with no question or answer content.
Consent
Consent under the personal data protection law. You accept the notice before you first use Ask LCA, and you can withdraw at any time.
Backup
An encrypted copy of the database, kept for about 14 days.
WAF (web application firewall)
A system that protects the website by filtering suspicious requests before they reach it. It is run by the Chiang Mai University team that manages the server.
API
The channel our program uses to send data to an AI provider for processing.
Credit
Each person's daily usage quota, counted from the actual amount of processing, to control the cost of the service.
Fallback
The backup path when the LLM cannot answer, for example when credit has run out: the system shows the FAQ and related links instead.
Golden set
A standard set of questions with the correct answers prepared in advance, used to measure quality the same way every round.
Evaluation (eval)
Running the golden set against the real system, then scoring every answer.
Gate
A minimum standard that the evaluation must pass before the service opens to members.
Privacy leak
An answer that discloses personal data that should not be disclosed.

Analytics cookies

This site uses Google Analytics to see aggregate usage — visitor counts and which pages get attention. It only runs if you accept.

No choice made yet